End-Gen_Technical_Description_C.pdf

UNCLASSIFIED

Table of Contents

1 Table of Contents

1.0 FUNDING OPPORTUNITY DESCRIPTION

1.1 Program Overview

1.1.1 Program Technical Challenges

1.1.2 Program Phases

1.2 Team Expertise

1.3 Program Scope and Limitations

1.4 Program Media, Modalities, and Data

1.4.1 Program Media

1.4.2 Program Modalities

1.4.3 Program Data (TA-2)

1.5 Testing & Evaluation (T&E)

1.5.1 Waveform Testbeds

1.5.2 Waveform Generation Hardware

1.5.3 Feasibility in Static Channel

1.5.4 Scalability in Dynamic Channels

1.6 Program Metrics

1.6.1 Waveform Objectives

1.6.2 When Is It a “Waveform”?

1.7 Program Waypoints, Milestones, and Deliverables

1.7.1 Waypoints

1.7.2 Milestones

1.7.3 Deliverables

1.7.4 Software Deliverable Formatting

1.8 Meeting and Travel Requirements

1.8.1 Technical Exchange Meetings (TEM) and Workshops

1.8.2 Site Visits

1.9 Period of Performance

1.10 Place of Performance

List of Tables Table 1: End-Gen Definitions Used Throughout the BAA Table 2: Nominal Test Set Conditions Throughout Program Phases for Both Technical Areas .. 12 Table 3: Non-Exhaustive List of Media Table 4: Common GAI Modalities Versus Likely End-Gen TA-2 Modalities Table 5. Interaction With End-Gen Frameworks Table 6: Non-Exhaustive List of Waveform Objectives Table 7: Primary Program Metrics by Phase and Technical Area Table 8. Recurring Waypoints Expected by All End-Gen Performers Table 9. Major Milestones (Events) Expected During the Life of the Program Table 10: Deliverables Expected from Performers Across the Life of the Program

List of Figures Figure 1: Notional Layers Defining the End-Gen Waveform Figure 2. Waveforms are the de facto atomic component of information exchange. End-Gen generative AI (GAI) frameworks generate waveforms in the wild directly (left, in situ) or in the lab (right, ex situ) and then transfer the waveforms into the wild Figure 3: End-Gen Program Design. Performer solutions generate outputs that are placed directly on radios (TA-1), or GAI is used in the lab with outputs placed on radios (TA-2) Figure 4: End-Gen Phase Objectives Figure 5: End-Gen Performer, T&E, and Government team key activities and milestones for the program. Dashed lines indicate joint events at T&E sites Figure 6: End-Gen frameworks consume static and dynamic inputs to generate waveforms appropriate for the channel Figure 7 T&E Infrastructure. The T&E team uses Performer frameworks to generate waveforms in the chosen scenarios. Scores for leaderboard scenarios will be posted to program wide leaderboard displays

INSTRUCTIONS

This Draft Broad Agency Announcement (BAA) is NOT a formal request for proposals.

Proposals are NOT being requested or evaluated at this time. A Final BAA will be released at a later date for which interested parties may submit proposals.

The Intelligence Advance Research Projects Activity (IARPA) is seeking interested parties to thoroughly review the Draft BAA Funding Opportunity Description and provide comments, questions, suggested changes, and feedback.

IARPA does not anticipate posting responses to any comments, questions, suggested changes, and/or feedback received; however, all input will be considered in developing the Final BAA.

Additionally, the Draft BAA Funding Opportunity Description is subject to change in the development of the Final BAA because of IARPA’s consideration of the input received from interested parties in response to the Draft BAA.

Interested parties will be given the opportunity to submit questions and comments in writing for IARPA response during Proposers’ Day and once the Final BAA is issued.

1.0 FUNDING OPPORTUNITY DESCRIPTION

The Intelligence Advanced Research Projects Activity (IARPA) often selects its research efforts through the Broad Agency Announcement (BAA) process. The use of a BAA solicitation allows a wide range of innovative ideas and concepts. The BAA shall appear first under Contract Opportunities on https://sam.gov, then the IARPA website at http://www.iarpa.gov/. The following information is for those wishing to respond to this Program BAA.

This BAA (IARPA-BAA-25-XX) is for the Endless Generative Waveforms (End-Gen) Program.

IARPA is seeking innovative solutions for the End-Gen Program in this BAA, which is envisioned to be a 42-month effort across all technical areas.

End-Gen lies at the crossroads of both communication, networking, and information theories (jointly referred to as simply “Comms” for the remainder of this BAA) and data-driven approaches to optimization sciences (simply referred to as “artificial intelligence (AI) and machine learning (ML)” [AI/ML] for the remainder of this BAA). End-Gen attempts to combine Comms with AI/ML to generate waveforms quickly and cost effectively. Since both fields, Comms and AI/ML, are robust and distinguished, there is invariably a crossover of the vocabulary used in either field.

Unfortunately, common words can have vastly different meanings depending on which field is being referenced. Due to this confusion, some common words, but with divergent meanings, are provided in Table 1 for consistency throughout the BAA. Words within the same field and variable meanings are also included in the table.

Table 1: End-Gen Definitions Used Throughout the BAA

BAA

Terms

End-Gen AI/ML Definition End-Gen Comms Definition

Channel A dimension on input samples or output products. For example, with image media/modality, the red, green, blue (RGB) color dimension of training data.

Multichannel outputs are also possible such as in complex-valued products.

Unfortunately, channel often has two meanings within the Comms community:

The “ether” between two radio nodes. Channel is different than simply the media as it includes other effects such as noise or other co-channel interferers in the same media.

A transmitter output or receiver input (e.g., one antenna “channel”)

Epoch If a training set is filled with a certain number of input samples then an epoch is one complete look at each of the labelled samples in the set.

A unit of time often used in Time Division Multiple Access (TDMA) channels. An epoch might be so many frames or subframes of data with a common temporal resolution.

Layer In the context of deep neural networks (DNN) and other AI/ML frameworks, a layer represents a single vertical slice of the network graph. For DNNs, the graph is made up of at least two layers.

The Open Systems Interconnect (OSI) model is fundamental to networks including the Internet. The most common OSI model includes seven different system blocks referred to as layers. For example, the lowest system block of the OSI model is called the physical (PHY) “layer.”

Media Generally, it refers to the objective inputs and outputs of the machine learning framework. Some common media for AI/ML problems are:

Text media such as in natural language processing machines

Image media, common to computer vision, object recognition, and security systems

Audio media for music classification or generation

Generally, it refers to the propagation media of the communications link over some physical channel. Common media are:

Radio Frequency (RF) is the most common wireless media often defined from a few kilohertz (kHz) to hundreds of gigahertz (GHz)

Optic media is much higher than RF frequencies into terahertz ranges

Acoustics or Ultrasonic media are most often in the kHz range

Modality This word will be used synonymously with Media and can be thought of as the data type that the AI/ML framework is interacting with. Common modalities include Text, Image, and Audio.

Comms modalities refer to modes that a link may operate in, such as with 5G, high rate, low latency, or high reliability modes.

Modes are defined by waveform objectives.

Network One type of graph used in AI/ML frameworks (e.g., neural “network”)

Multiple radio nodes communicating with each other using waveforms for information exchange

Sample A single example from the training dataset drawn upon to train the network. For text media a sample might be a single page of words. For image media a sample might be a single picture or object.

A single unit of the waveform output from the hardware (HW) frontend. A sample is a discrete, but analog, value of the waveform. Samples are often complex values to include phase and amplitude content of the waveform.

Train The process of cycling through multiple epochs of the training set in order to update AI/ML weights to achieve objective(s).

In order to obtain Channel State Information (CSI) of the channel transfer function, pilots are often used to train the channel.

Radio “Radio” will not be used exclusively for wireless RF hardware as is common. Radio will refer to any hardware needed to convert samples into energy appropriate for the environment. Radios, as defined in this BAA, are not constrained to RF frequencies or other electromagnetics. An appropriate digital-to-analog converter (DAC) plus transducer (e.g. speaker) for the acoustic channel is a “Radio”; hardware with a light emitting diode (LED) and photodiode is a “Radio”; and so on.

Waveform While not generally used in the AI/ML community, a waveform does have multiple confusing definitions in the Comms community:

A waveform is the energy propagating through whatever media between two nodes. A pulse is a waveform.

A waveform is the end-to-end protocol for a communications link. Bluetooth is a waveform.

This BAA will not attempt to disambiguate the usage of the above words through any special formatting. When a shared word is used in this BAA the reader is encouraged to carefully consider the usage and context of the passage should always be considered.

What is a waveform?

There is additional confusion on defining “waveform” even within the Comms community. Some usages of the term refer to the actual energy physically propagating through the environment whether via electromagnetic fields, acoustic waves, or even exotic waveforms such as temperature gradients. In this case, the waveform is simply the time-energy relationship connecting two different nodes that allows for information exchange. Unfortunately, especially in government terminology, “waveform” also refers to the entire protocol stack of communications. For example, Link 16 is called a “waveform” while the modulation output, a subblock within Link 16, is also called a “waveform.” Waveform will be used in both definitions throughout the BAA and the reader is encouraged to take care from context in determining which definition is appropriate.

End-Gen is focused on frameworks capable of generative waveforms. Which definition is a waveform then in the context of the program? Figure 1 shows how these two definitions play out within the context of the End-Gen program. End-Gen waveforms must command and control the radio (Layers 2 and 3) while turning bits into PHY layer samples (Layer 1) and meeting real-world objectives. Offeror proposals should show no confusion on the goal of the End-Gen program and what is being generated. A generative waveform is all the processing needed for the lower three layers, as shown in Figure 1, not just the PHY layer output samples.

Figure 1: Notional Layers Defining the End-Gen Waveform

Network (NET) Layer In the OSI model, the network layer is responsible for facilitating data transfer between two different networks. The network layer normally receives data from higher OSI model layers and subdivides the data into packets for transmission. End-Gen will not interface with higher model layers or model transmission protocols normally associated with the higher layers of the OSI model. Instead, End-Gen Test and Evaluation (T&E) and Performer teams will access scripted sources of data that emulate mission goals and create measurable performance metrics of each phase of the End-Gen program.

Medium Access Control (MAC) Layer The MAC layer is like the network layer, except the MAC layer facilitates data transfer between two devices on the same network by managing channel usage. The MAC layer takes packets from the network layer and breaks them into smaller pieces called frames. These frames are then intelligently transmitted over the medium. The MAC layer is responsible for flow control and error control in intra-network communication. The End-Gen solution will also need to show intelligence when accessing the channel.

Physical (PHY) Layer This layer includes the physical equipment or medium involved in the data transfer, such as the cables, antenna, speakers, or other transducers. At the receiver, this is also the layer where the noisy waveform is received and gets converted back into a bit stream. The physical layer of both devices must also agree on a signal convention so that bits transmitted correspond to bits received.

In short, in the End-Gen program, a waveform is a protocol that decides 1) Who to transmit to (NET layer), 2) When (time), Where (frequency, space) to transmit (MAC layer), and 3) What to transmit (PHY layer) resulting in a waveform in the real world. Higher layers will not be part of the End-Gen program. Note that Offeror solutions are not the waveform-in-space itself; Offeror solutions are the tools, mechanisms, and processing - the how - of waveform generation between radios.

What is a radio?

Generative waveforms will be created on radios in the End-Gen program. Since a generative waveform in its most basic form is a mapping of bits to waveform samples, it can be generated in any environment that human created waveforms function in. Throughout the BAA, the hardware used to transmit and receive waveforms will be generally referred to as “radios.” A universal software radio peripheral (USRP) connected to a desktop computer and appropriate antennas will be called “radio.” A laptop computer with a microphone/speaker combination for acoustic waveforms will also be called a “radio.” In short, radios are anything that are capable of creating generative waveforms and pushing them into the environment.

1.1 Program Overview

National security missions need to generate, store, use, transmit, and receive information and data both in secure facilities and “in the wild.” Waveforms are often used as the method for the transmission and reception of the data generated in these types of missions. In general, a waveform is any method used to pass information between two entities by transducing physical energy, that represents the information being shared, into the physical world environment. As shown in Figure 2, the End-Gen program is interested in two methods of generative AI (GAI): generating novel waveforms in the wild directly or generating in a lab and then transferring to the wild.

Figure 2. Waveforms are the de facto atomic component of information exchange. End-Gen generative AI (GAI) frameworks generate waveforms in the wild directly (left, in situ) or in the lab (right, ex situ) and then transfer the waveforms into the wild.

The prevalence of waveform design is shown by the significant commercial private sector investment to create waveforms that satisfy the population’s need for throughput and reliability of information. The popularity of connected humans has led to a ubiquitous waveform vocabulary such as Wi-Fi, 4G/5G, and Bluetooth. While U.S. government (USG) waveform names are lesser known to the general population, USG has also invested heavily in waveform development for their own mission needs. Due to the broad nature of missions, waveform development can be slow, expensive, and repetitious to meet needs in a changing world. A waveform development and acquisition cycle may take years to field new waveforms, potentially making them obsolete before becoming fully operational.

1.1.1 Program Technical Challenges

End-Gen's ambitious goal is to change the paradigm of waveform development. If successful, End- Gen will support, enable, and augment numerous intelligence community (IC) mission requirements by creating a time- and cost-saving approach to robust information exchange – the generative waveform. End-Gen will completely alter the approach to developing solutions for new wireless capabilities. The End-Gen approach removes the human from problem solving roadblocks (how to create the right waveform) and moves humans toward problem creation efforts (how to best complete the mission). In other words, humans focus on the problems they are interested in and let the machine determine new algorithms, new waveforms – even new inventions – that will satisfy IC needs. Comms designers can focus on what is interesting and meaningful and let End- Gen do the rest.

Thus, End-Gen proposes a program seeking to develop novel methods of arbitrary generative waveforms, enabling the intelligence community, and wider USG, to more easily and cheaply create wireless information exchange protocols without costly, delayed human development cycles. This will be accomplished across two developmental technical areas (TA):

TA-1 In Situ Discovery: End-Gen generative frameworks, loaded onto smart radios1 are placed within the desired environment, given human mission objectives, and generate waveforms through an adaptive training method.

TA-2 Ex Situ Conservation: End-Gen foundational2 generative frameworks, prompted to create waveforms suitable for the desired environment and human objectives, create outputs which are placed onto smart radios.

The objective is for End-Gen to generate waveforms that satisfy mission needs. Rather than humans spending time developing algorithms that satisfy mission needs, humans focus on the mission and objectives themselves. The goal will be to generate waveforms either in situ (TA-1) within the environment itself, or ex situ (TA-2) within the laboratory, and then transfer these waveforms to the environment for testing. Each TA poses unique and shared challenges to make generative waveforms a reality.

In Situ Discovery Challenges

TA-1’s name of In Situ Discovery reflects the goal of “discovering” the waveform knowledge in the environment itself both real-time and just-in-time. In its most basic form, nodes equipped with End-Gen TA-1 frameworks have no fixed ability to communicate with each other. There is no side channel or any other connection that allows the exchange of information. Offerors must detail in their proposal how to use epochs of training for these radios to create waveforms that allow communication. End-Gen is seeking an end-to-end waveform generation solution, where input bits are turned into appropriate samples that will be sent over the channel. The system needs to automatically perform this mapping while output samples, when sent over-the-air (OTA3), meet the objectives defined by the program and given to the solution framework in the moment. The End-Gen in situ objective is to create information-bearing waveforms without any prior knowledge of the environment and only immediate knowledge of the waveform objective.

End-Gen TA-1 presents several significant challenges. Whatever training is needed to generate the waveform must occur over the same imperfect channel that the radio nodes are in. A fundamental premise of the End-Gen program is that waveforms are generative on-the-fly as needed. This presupposes that no feedback channel, even a low-rate feedback channel, is available, or should be proposed, in the Performer’s solution. Any energy transmitted from any node is subject to both the environment the nodes are placed in and any metric scoring performed by T&E. While there will certainly be a “train” period and a “test” period, any emanations during training will be held to the same environment as during testing. Additionally, the End-Gen solution framework processing requirements must fit within the hardware processing capability. Solutions like large generative models will be inappropriate for In Situ Discovery as hardware sizes shrink throughout

1 See Table 1. A “radio” refers to any hardware capable of energy transduction between two points.

2 Not to be confused with Foundation Models which are likely out-of-scope.

3 OTA means “not simulated” here. Both subsurface and wired waveforms are “OTA” when emitted.

the life of the program.

Ex Situ Conservation Challenges TA-2’s name of Ex Situ Conservation reflects the lofty goal of creating a foundational model containing all knowledge related to waveform creation. Unlike TA-1, which generates waveforms directly from input bits in the environment itself, TA-2 will generate all the building blocks needed for a smart radio to create waveforms once emplaced. A TA-2 End-Gen framework is an oracle that is approachable and promptable. Through a series of prompts, End-Gen will produce outputs, which after Lexicon translation, can be placed on the radios and generate the needed waveforms.

Where TA-1 frameworks are hardware constrained, there is no constraint on TA-2 frameworks except for excessive financial cost that may be out-of-scope. It is the TA-2 framework outputs that will generate the waveform and are constrained on the hardware similar to TA-1.

End-Gen TA-2 presents several significant challenges. Where TA-1 solutions are generally predicted to be unexplainable and monolithic in nature, a direct mapping from bits to samples, TA- 2 has the added difficulty of explainability. For example, a sequence of prompts will lead to a single signal processing or computation block which process is then repeated to create a series of computational blocks. Each TA-2 output is clearly accomplishing some given processing task.

Thus, unlike TA-1, TA-2 outputs are required to be explainable and training datasets become a major portion of final solutions. Offerors will be required to collect and annotate research and development (R&D) datasets throughout the program lifecycle and share them across the program with T&E. All data collections must not include Human Subject Research (HSR) or require Institutional Review Board (IRB) approval.

Generative Waveform Challenges

The basic program design is depicted in Figure 3. A diverse set of objectives are provided to radios to satisfy some particular mission. Hardware is available to transduce energy in some particular environment. Offerors to the End-Gen program shall propose solutions for either the in situ (TA-

1) or ex situ (TA-2) approach to generative waveforms. TA-1 solutions are placed directly on the radio. TA-2 solution outputs, after translation through some Lexicon, are placed directly on the radios. These solutions need to consume the given objectives and function in the unknown environment, all while meeting the program metrics. Multiple offers from the same team are allowed for the two separate TAs but must come in separate submission packages.

Figure 3: End-Gen Program Design. Performer solutions generate outputs that are placed directly on radios (TA-1), or GAI is used in the lab with outputs placed on radios (TA-2).

1.1.2 Program Phases

End-Gen is proposed as a 42-month research program consisting of one 12-month ramp-up Phase 1, one 18-month deep dive Phase 2, and the program completed with one 12-month integration Phase 3 ensuring clean transition. Phase 0 for T&E will focus on creating all the needed interactions between future Performers and the testbed including application programming interfaces (APIs), Lexicons, and doctrines that need to be followed. In TA-1, Performers will need to demonstrate increasingly sophisticated generated waveforms as defined by increased objective count and multiple media. For example, while Phase 1 may only require one medium (e.g., RF) and three objectives such as bit error rate (BER), throughput, and bandwidth, Phase 3 will require multiple media (e.g., RF plus powerline and/or acoustics) and add objectives like latency and medium access control. TA-2 will follow similar increasing phase difficulty as TA-1 with some additions. Phase 1 for TA-2 prompting will come exclusively from Performer teams. Phase 2 will stress End-Gen generality by solving problems from government team prompts. Finally, Phase 3 will have End-Gen inventing/discovering basic algorithms that do not yet exist faster than human generated algorithms through 3rd party prompting.

The progression of objectives for both TAs across the program phases is shown in Figure 4. As generative waveforms are a relatively new technology, the first phase is focused on feasibility. The program must show that GAI can be applied to waveforms. Infeasible solutions will not move forward. After bringing waveforms to life in Phase 1, the solutions must scale appropriately to be useful in live environments and missions. The program will end in Phase 3 as End-Gen generates waveforms for transition teams to use in the field.

Figure 4: End-Gen Phase Objectives

Offerors are required to propose to all phases (Phases 1, 2, & 3); however, Offerors can submit to either, or both, Technical Areas (TA-1 & TA-2) of the Program under this BAA. Any Offeror proposing to both TAs should submit separate proposal packages. Proposals that omit any of the phases or propose a solution outside one of the TAs, will be considered non-compliant.

Test set conditions will expand throughout the program phases to increase the number of objectives required for each waveform, increase the number of node communications within the same network, increase confounders and other channel effects, and lower the available power or receive signal to noise ratio (SNR). A listing of nominal, and exemplar, test-set conditions throughout the phases is provided in Table 2. The exact conditions are subject to potential update during program execution.

Table 2: Nominal Test Set Conditions Throughout Program Phases for Both Technical Areas Phase 1 Phase 2 Phase 3 Radio # Nodes 2 3 4 Transmit Power < 1 W (for RF) all phases Hardware Processing Server Desktop Embedded Waveform Coherence Time > 5 minutes all phases # Antenna (Channels) 1 1 2+ Environ Channel Effects None Mild Severe Channel Coherence Time Infinite O(Minutes) O(Seconds) # Different Media 1 2 3+

Testing conditions throughout the program phases can be roughly broken into two categories.

Radio conditions refer to internal hardware that the Performer directly controls and solutions must self-compensate to work with. Environ conditions are external to Performer hardware, and therefore out of their control, but certainly affect what waveform will be generated and must be handled accordingly.

Radio Conditions

In general, Performers will have control over a predefined set of radios during a test event to an extent. Some parameters, such as center frequency, may be constrained for ease of testing. As the number of radios increases throughout the phases, the End-Gen framework must exist on all radios and be able to create appropriate data links between each. Transmit power will be capped at 1 W (for RF scenarios) for all phases. The time at which waveform objectives change, called Waveform Coherence Time, is capped at 5 minutes meaning objectives remain static for at least that long.

Near the end of the program, additional antenna will become available and may be utilized to improve waveform performance.

Processing power is an important constraint for the program. Hypothetically, the intelligence of the generative solutions is most likely tied to the available processing power of the nodes. The hardware/processor versus intelligence/performance trade-off is important for mission partner understanding. To help answer this question, performer solutions will either be placed on variable sized hardware or virtually shrunk by limiting available access to compute resources.

Environmental Conditions

The communications channel will become increasingly difficult throughout the phases of the program. In Phase 1, Performers can expect a fairly static channel in only one type of media. By the end of the program, waveforms will need to be generated in a variety of channels and overcome significant co-channel effects. Performers will not be told before test events which set of channel conditions or media will be used for testing and scoring. Intelligence should come from the machine and not the human team. End-Gen frameworks, the software itself, will be told immediately before testing and must accommodate whatever channel effects are used in-the-moment.

Leaderboards are planned to provide ongoing measurement of performance with rankings for all Performers. IARPA will also provide access to self-test scenarios for Performers to run algorithms on their own hardware with results submitted to a T&E hosted analysis for scoring and Leaderboard posting. Monthly Leaderboard submissions (initially to the self-test leaderboard, and subsequently to the sequestered T&E leaderboard) will be required to provide ongoing insight into Performer progress. Semi-annual waypoints will be incorporated where performance will be reported to IARPA management and software will be delivered to T&E to facilitate independent builds and runs from source code. Performer data collections and annotations must be shared with other Performers, as filtered through the T&E teams, after each semi-annual performance reporting waypoint. Key program activities and milestones are shown in Figure 5.

Figure 5: End-Gen Performer, T&E, and Government team key activities and milestones for the program. Dashed lines indicate joint events at T&E sites.

For TA-2, Performer teams will be responsible for obtaining (collecting, curating, simulating) and annotating data in support of proposed R&D objectives on their own. The government will not collect, curate, or annotate any data for evaluation; however, the government reserves the right to share public training sets, or private training sets created within the program itself, with all Performers. Performers will submit containerized (nominally Docker) software using/conforming to the framework and APIs specified by the T&E team for independent runs on sequestered OTA scenarios.

In addition to compatibility with the End-Gen T&E API, T&E must also be able to interact autonomously with the Performer solutions through the End-Gen Lexicon. Where the API allows Performers to control radio HW on the testbed, the Lexicon allows the government team to interact with Performer solutions. This interaction includes methodology to feed objectives and prompts into the Performer solution. Unique to TA-2, the Lexicon allows T&E to interpret End-Gen solutions outputs onto the radios in a consistent manner. For example, in its most basic form, the Lexicon may simply be a particular program language that enacts inputs/outputs of signal processing blocks.

Phase 1 – Focus on Creation of a Generative Waveform

During the initial 12-month Phase 1, End-Gen Performers will research and develop the capability and demonstrate the feasibility for generative waveforms in relatively static environments.

Waveforms will truly be generated by the machine. Starting after the first quarter, monthly submissions will be required to run against the leaderboard scenario(s) throughout the program (scoring is against both the self-test leaderboard and the sequestered T&E leaderboard). Runtimes will not be fixed and only long enough to generate a statistically significant number of bits to calculate program metrics. In general, scenarios will be no longer than five minutes for RF scenarios but may be longer for lower capacity channels when run on the T&E hardware for sequestered dataset runs.

During this phase, each scenario will be run in a relatively static environment free from excessive confounders such as co-channel interference and jamming. Additionally, the effective receive SNR will be high compared to nominal operating conditions. However, to encourage robustness in generated waveforms, a single static channel effect may be added to discourage trivial, human generated, solutions. Multiple channel types will be included, but additional objectives beyond the minimal needed to be stamped “waveform” will not be included. Performer systems will be required to generate waveforms ad hoc and in situ.

Phase 2 – Focus on Scaling Generative Waveforms

Phase 1 will have shown that the machine can generate a waveform in static, well-known channels without human intervention. Phase 2 will show that Performer solutions can scale to more meaningful channels. The number of nodes equipped with End-Gen frameworks will increase. The size of the hardware will scale down. The available power will scale down. T&E scenarios will increase in complexity with more robust and richer channel environments. Rather than static additive white Gaussian noise (AWGN), additional channel occupancy will occur from other sources. Power may be more limited to the radios. The End-Gen nodes may need to demonstrate command and control over the radios to make transmission decisions such as medium access (Layer 2). Where training time was open ended in TA-1 Phase 1, in TA-1 Phase 2 Performers will be scored against each other on how long it takes for their nodes to generate a waveform. Phase 2 will have a duration of 18 months.

Phase 3 – Focus on Transitioning Generative Waveforms

Phase 1 and Phase 2 will show that waveforms can be generated by the machine and that solutions can scale toward practical values that are meaningful to the community. Besides intelligence in waveform generation, the End-Gen solutions also show intelligence in higher layers of the OSI stack by determining when and who to transmit waveforms to. With feasibility and scalability of generative waveforms determined, the program can then focus on real world problems from transition partners. For TA-1, hardware constraints and environmental effects will be closer to real mission needs. For TA-2, machine prompting must be more approachable as transition partners are allowed to interact with the End-Gen solutions and create exemplar waveforms based on their interests and objectives. During Phase 3, the waveform framework should be mature enough to be generated, frozen, and then used in actual environments. Radio and environmental conditions will be added to ensure this is true.

1.2 Team Expertise

End-Gen lies at the crossroads of AI/ML and Comms. Collaborative efforts and teaming across candidate Offerors are highly encouraged. It is anticipated that teams will be multidisciplinary and may include expertise in one or more of the disciplines listed below. This list is included only to provide notional potential expertise areas for the Offerors; satisfying all the areas of technical expertise below is not a requirement for selection and unconventional or innovative team expertise may be needed based on the proposed research. Proposals shall include a description and the mix of skills and staffing that the Offeror determines will be necessary to carry out the proposed research and achieve program metrics.

Artificial Intelligence Deep Learning Machine Learning Communications Propagation Networking Information Theory Large X Models Wireless Protocol Stack Software Defined Radios Modeling and Simulation Software Engineering Software Integration Systems Integration Continuous Integration / Continuous Delivery Other (describe)

When choosing team skillsets, keep in mind Performers are required to report monthly progress using a similar metric score as calculated by T&E. Ideally these scores will be similar as found during T&E test events; however, no effort will be made to determine one-to-one performance.

Therefore, while some scenario generation capability is expected by Performer teams it is not expected to match the sophistication of the T&E testbed. Thus, some team expertise is expected on scenario generation and channel analysis but should not consume an excessive amount of team composure.

1.3 Program Scope and Limitations

At its core, the End-Gen program is about making diverse waveforms quickly and cost effectively.

And, truly, the stretch goal of the program is creating a framework capable of generating all meaningful waveforms that have been, or might be, invented by humans to handle mission needs.

While solutions are expected to be complex structures, the underlying need is straightforward:

generate waveforms. The Performer’s solutions, once placed on radio hardware, must consume two types of inputs: dynamic and static. Dynamic inputs are unknown beforehand; these include information bearing bits, that are the primary reason for waveform creation, and observations of the channel conditions. Static inputs, within the coherence time of the waveform, include objectives of the desired waveform which are fed into TA-1 as mathematical functions and via prompts for TA-2. Figure 6 shows the general structure of the solution for all proposals.

Figure 6: End-Gen frameworks consume static and dynamic inputs to generate waveforms appropriate for the channel.

It is extremely important to note the differences in each technical area for how the output “waveform” is created. For TA-1, Offeror solutions must perform all calculations needed on dynamic and static inputs to produce a string of complex samples at the output which make up the waveform. For TA-2, Offeror solutions must create the signal processing blocks needed, that operate on the dynamic source of bits, to produce a waveform. In TA-1, the solution is placed directly on the radio in the environment. In TA-2, the framework prompted outputs are placed onto the radios and then placed in the environment. Thus, the outputs of the TA-2 solutions are what must consume the bit stream and channel observations, while TA-1 frameworks consume bits directly.

Proposals shall explicitly address:

Underlying theory: Proposed strategies to meet program-specified metrics must have firm theoretical bases that are described with enough detail that reviewers will be able to assess the viability of the approaches. Proposals shall properly describe and reference previous work upon which their approach is founded. Reporting on benchtop demonstrations is encouraged. Demonstrations of atypical generative frameworks may also be reported even if outside the Comms modalities.

Research & Development approach: Proposals shall describe the technical approach to meeting program objectives and metrics. The technical approach must clearly demonstrate how the series of operations can transform bits into waveform samples.

Technical risks: Proposals shall identify technical risks and proposed mitigation strategies for each identified risk.

Software development: Proposals shall describe the approach to software architecture, modularization, and integration including how continuous development and integration will include the T&E teams.

The following areas of research are out-of-scope for the End-Gen program:

Research relying on pretrained Large Language Models whether public or proprietary Foundation Models that require excessive costs in input datasets or training Research that does not have strong theoretical and experimental foundations Approaches that rely solely on integration of existing technologies without experimental or scientific improvements Development of hardware or sensors Approaches that consist merely of integrating currently existing software systems Research of non-generative waveforms Research on modern human-derived signal processing Methods that require a human in the loop Methods that require fixed algorithms in the loop Disproportionate focus on MAC/NET layers instead of PHY

To reiterate, while the term “AI/ML” is used throughout the BAA for simplification, solutions that are not specifically focused on AI/ML are certainly allowed. While AI/ML methods are commonly expected, non-AI/ML methods are also welcome; however, any submitted method should generate waveforms on-the-fly and not resort to fixed allocation of preexisting waveforms which may not cover the breadth of testing, or any of the other out-of-scope research areas, unless Offerors can clearly articulate why such approaches will meet all program objectives in any channel media.

As part of TA-2 scope, performers are expected to use data from external sources provided that the data is consistent with an approved privacy plan, which may be shared across all Performer and T&E teams, and is not proprietary. Please include a list of any datasets that you wish to share within your proposal for the effort. Any data used for training shall be made available to all selected Performer and T&E teams on End-Gen and be provided with Government Purpose Rights (GPR).

Datasets need not be made public beyond the program teams.

Delivered solutions must include all resources and execute autonomously, not relying on internet access during execution or on any proprietary modules. There may also be the ability to accumulate external data for system reference. If a solution requires external data sources, proposals should describe the necessary data and what would be required to obtain such data for offline use.

1.4 Program Media, Modalities, and Data

Achieving program metrics will require Performers to develop methods for command and control of hardware, for waveform generation in a wide variety of channels, and for handling a variety of objectives. Channel media will range from high frequency RF channels to low bandwidth acoustics. Unlike traditional GAI that focus on single modalities such as natural language or images, End-Gen GAI frameworks must cover a range of signal processing and mathematical modalities to cover all knowledge and processing needed for waveforms. This requires Offerors to carefully consider program media, modalities, and data.

1.4.1 Program Media

In general, human waveform designers need a firm understanding of the media and channel conditions to achieve objectives of the waveform itself. The primary purpose of a waveform is to move data through whatever channel it is placed in while maintaining a certain set of observables.

Since the channel is unknown to Offerors prior to testing, End-Gen is seeking solutions that are agnostic to the media. Once an End-Gen framework is introduced into the media, either directly placed into the media as with TA-1, or indirectly informed of the media through prompting in TA- 2, the waveform generated should be appropriate for that media. For example, a valid End-Gen framework would never generate a waveform with a 1 GHz bandwidth in the acoustic channel as this has unrealistic propagation expectations.

Table 3: Non-Exhaustive List of Media

Media Properties Acoustic Acoustic channels are roughly 100-20,000 Hz. Transducers can be common hardware such as speaker and microphone combinations.

Exquisite Any transfer of energy can technically be modulated and result in an exchange of information. Vibrations between two points, flickering lights, or temperature gradients are all possible media for communications. As Offeror solutions mature, the program will introduce new media to validate the general nature of End-Gen generative waveforms.

Radio Frequency

(RF)

RF environments are generally wireless and nominally range from a few hundred MHz to several GHz center frequencies. Bandwidths are expected to be in the hundreds of kHz to tens of MHz wide.

Ultrasonic Ultrasonic channels range beyond hearing levels or anything over 20 kHz frequencies. Ultrasonic waveforms are transduced with specialized hardware.

Wired Wired environments have a direct connection between transmit and receive channels using an appropriate conductor. A coax cable with an attenuator connecting two ports on the radio is a wired channel.

Given End-Gen’s ability to adapt to any media presented, an exhaustive list of media prepared for the program will not be presented here. Offerors should develop frameworks capable of any possible media and solutions will only know at the time of testing what the media will be for that particular scenario. However, common media have a greater likelihood of being used in scenarios as they have a greater impact on community usage. While several media are planned throughout the program, at program kick-off the RF channel is the only media Performers will be told will be scored. Even so, Offerors proposals must report on how their frameworks will work with any channel media presented in the T&E labs.

1.4.2 Program Modalities

GAI have become ubiquitous. Modern uses of GAI are commonly defined by two segments of their framework: input/output modalities and the machine optimization goal. Single, or multiple, input modalities define what domain the AI/ML is working on, with the most recognizable domains being: text, image, or audio (TIA); single, or multiple, AI/ML outputs can be in that same, or perhaps different, modality. The AI/ML machine goal itself is often further distinguished by two types: a discriminative AI attempts to find boundaries within some input modality while a generative AI attempts to understand the entire distribution of that modality. And while discriminative AIs may perform better at class labeling, generative AIs have a more robust knowledge of the entire input distribution and can create original samples from the input mode.

GAIs, or any similar optimization framework, with uncommon modalities, that can create original samples in those modalities, is at the heart of End-Gen. The GAI modality cannot be trivialized in proposals.

Table 4: Common GAI Modalities Versus Likely End-Gen TA-2 Modalities

Common GAI Modalities End-Gen GAI Modalities Audio such as music generation or text-to-speech generation.

Algorithms as a basic modality define a sequence of steps on inputs to achieve some objective.

Image such as computer vision or picture generation.

Math modalities focus on e.g. calculus, statistics, and physics generation (not solvers).

Symbolic such as might be found in GAI frameworks used for creating coding languages like Python or C code.

Samples are a representation of a discretized waveform. The sample as a modality is a viable approach in GAI creation.

Text (or Language) such as natural language processing and content classification.

Signal Processing in its most basic mode is concerned with taking inputs in a certain data type and output in another data type.

Text, Image, and Audio Modalities (TA-2)

GAI has seen significant research and development recently as it is being applied to many problems. Text-to-text generative AI, such as ChatGPT, has taken on a nearly ubiquitous spot in the population’s lexicon with its ability to help human users generate useful snippets of text for their own purposes. Text-to-image generative AI, such as DALL-E, was created recently to generate unique images, using a similar transformer model as ChatGPT, but with output data images rather than natural language text. Image-to-text generation, often called image captioning, is used to produce natural language descriptions of input images. There is also GAI for generative audio such as with text-to-speech synthesis.

TIA modalities with GAI have seen great application so far and can certainly help scientists and engineers perform better in their fields; however, TIA modalities may not have the robustness necessary to generate waveforms. The use of TIA modalities for this purpose may be hindered by their ability to capture the meaning of scientific or engineering topics to make the discoveries necessary to satisfy waveform objectives. Furthermore, TIA AI/ML often suffers from opinion, bias, and falsehoods while the modality used for waveform generation must only be true. If Offerors propose using TIA modalities for generative waveforms, they must be very clear on how generated TIA can translate into a waveform that satisfies program metrics. Different modalities other than TIA, such as: math, graphs, signals, or symbolic inputs/outputs may be necessary to create generative waveforms.

Math, Algorithms, and Signal Processing Modalities (TA-2)

Less common GAI modalities may have greater success with generative waveforms. These might include math, algorithms, and signal processing (MASP) modalities. An End-Gen foundational model, when fully realized, should have the opportunity to create marginal advances to MASP problems by generating novel MASP products regardless of the modality chosen. These marginal improvements, fed back into future End-Gen machine iterations, in a positive feedback fashion throughout the life of the program, may provide improvements to the broad field of waveform design.

Math: Generative math is created by an AI/ML framework that is trained on, and consumes, TIA, MASP, or other symbolic inputs to generate outputs in the math modality: theorems, lemmas, equations, etc. For this program, these products may be preexisting human-discovered math knowledge or new solutions found for particular End-Gen problem sets. Generative math is not a “math solver” itself; rather, generative math will take appropriate inputs to create something novel which helps create the waveform itself.

Algorithms: Generative algorithms are created from an AI/ML framework that is trained on, and consumes, TIA, MASP, or other symbolic inputs to generate an algorithm, or sequence of: steps, variables, feedback loops, updates, etc. to perform some action, based on a given objective function. A generative algorithm AI/ML does not perform steps of an algorithm itself; rather, it learns from the symbolic input modality how to arrange the steps of the algorithm, the truth behind each step, and objective optimization. An algorithm that performs a sequence of steps on samples, with the final goal being a sampled waveform, would be applicable to End-Gen.

Signal Processing: Generative signal processing is created from an AI/ML framework that is trained on, and consumes, TIA, MASP, or other symbolic inputs to generate signal processing products useful to an engineering discipline: filtering, modulations, transforms, waveforms, etc.

For example, generative signal processing may create a known error correction coding scheme, after being trained on topics such as bit manipulation and additive noise. The generative signal processor should output a signal processing function the human can use in creating the generative waveform.

While MASP modalities have an obvious correlation with waveform…