Two to ten bits per second of intent, expanded by context — bridging the bandwidth mismatch between biosignals and complex control
The National Science Foundation awarded $1.25 million to IN VIRTUALIS to develop AI middleware that transforms minimal biosignals into context-aware, multi-step actions. Current brain-computer interface and electromyographic systems typically provide only two to ten bits per second of control information, insufficient for precise task execution.
Grant overview (primary data)
- Award amount$1,250,000
- RecipientIN VIRTUALIS LLC (Washington)
- ProgramSBIR Phase II
- Period2026-08-15 〜 2028-07-31
- FunderU.S. National Science Foundation (NSF) / NSF
Key points
- AI middleware that transforms minimal biosignals into context-aware, multi-step actions ($1,250,000).
- Current brain-computer interface and electromyographic systems provide only two to ten bits per second, insufficient for precise task execution.
- Three innovations: real-time affordance prediction from video and gaze, hierarchical action decomposition, and adaptive user modeling through reinforcement learning.
- Targets are sub-500-millisecond end-to-end latency, at least 85 percent intent prediction accuracy, and interoperability across hardware platforms.
- The program is SBIR Phase II, awarded to companies; seven of the awards this site holds as of 2026-09-02 fall under it.
1Less information available than is needed
Reading intent from brain activity or muscle signal has advanced. What this project points out is that the sheer quantity of information available is small: current systems deliver on the order of two to ten bits per second — less than one character of text per second. Controlling everyday actions precisely requires far more. The project calls that gap a bandwidth mismatch.
2Making up the difference with context
- 1Affordance predictionInfers in real time, from video and gaze data, which actions are contextually appropriate in the scene
- 2Hierarchical action decompositionMaps simple user inputs to multi-step behavioral sequences
- 3Adaptive user modelingPersonalizes control strategies through reinforcement learning
- 4TargetsSub-500-millisecond end-to-end latency and intent prediction accuracy of at least 85 percent
The idea is not to amplify the sparse signal coming from the user but to narrow, from the surrounding situation, what action would be meaningful here and now, so that a small signal suffices to choose among the remainder. If a cup is in front of the user and their gaze is on it, the candidates are few. Once context has reduced them, two to ten bits is enough.
3Middleware as a position
What the project builds is not a finished device but a layer between devices. For device manufacturers, a standardized, interoperable solution reduces in-house engineering cost and speeds deployment.
Commercially, the strategy leverages runtime licensing to original equipment manufacturers, supporting adoption through existing distribution channels, with initial deployments targeting specific manufacturers serving research and clinical institutions.
4The SBIR framework
The program is SBIR Phase II, awarded to companies rather than universities and designed to carry Phase I results toward commercialization. Of the NSF awards this site holds as of 2026-09-02, seven are SBIR Phase II, and companies appear among the 93 institutions. Funding for basic research and funding premised on commercialization sit under the same foundation. Amounts are the obligated amount as of the check date and may change.
Why it matters
Making a sparse signal effective by reducing the choices rather than amplifying the signal transfers to any interaction where input is constrained. Offering the layer between devices as a standard, rather than the device itself, is also worth noting as a business shape.
FAQ
What is the bandwidth mismatch?
How is the difference made up?
Sources (primary)
Source: NSF Award Search (U.S. National Science Foundation, public domain). Amounts are the obligated amount. For privacy, we do not handle principal investigator names.
- NSF Award (original, official)
- NSF Award ID: 2545788