Summary
“Physical AI” names the stack that connects learned models to actuators: sensors that capture the world (cameras, event cameras, tactile skins, inertial and gaze signals), a learned policy that maps perception plus a language instruction to motor commands, and the compute that has to run that policy on a moving, battery-powered platform. The dominant policy architecture in the current literature is the vision-language-action (VLA) model, and within that the diffusion-based variants, which give the best motion quality and generalisation but must be evaluated at control frequencies of roughly 50-200 Hz, imposing hard latency and energy budgets at the edge. A parallel line of work replaces or augments the reactive policy with a world model: a learned predictive core that compresses observations into states and simulates action-conditioned futures for planning.
The research frontier is currently about the failure modes rather than the demonstrations. VLA perception is fragile at the sensing stage: under 20 lux, an image-only policy scored 0% on a pick-and-place task, rising to 60% when accumulated event-camera maps were simply overlaid on the RGB frames. Neuromorphic sensing is being pushed for the same reason on drones, cars and underwater vehicles, where motion blur and harsh light break frame-based imaging. On the compute side, the work is co-design: exploiting temporal similarity between consecutive control inputs to skip redundant bit-level operations, or, on AR glasses, using gaze, pose and inertial cues to keep only the informative parts of egocentric video for a 27.5x memory and 24.3x energy reduction versus a full-video baseline.
Evidence base
Open questions
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.