Expert motor intent, carried into autonomous machines
RAST infers what skilled operators optimize for when they move, then computes actions from that intent as bodies, payloads and physics change. We are applying it to autonomous drones and robotic surgery.
Under the hood: inverse optimal control to infer the objective, optimal feedback control to act on it, and deterministic safety limits on every command.
Skill breaks when the dynamics change
A drone picks up a payload and its inertia changes mid-flight. The wind shifts. A new airframe handles differently. A surgical robot moves to a new instrument or a new task.
Behavior that was demonstrated or trained under one set of dynamics is brittle under the next. Expert operators adapt without thinking. Their machines do not.
The missing piece is the intent behind the skill: what the expert was optimizing for. Intent is the part worth transferring.
RAST, the execution layer
RAST runs alongside an existing control stack. It turns expert demonstrations into a compact intent package, then computes actions from live state and that intent, inside hard safety limits.
Offline tooling that infers the objective behind expert behavior, separates stable traits from context-dependent adjustments, and exports a deployable intent package.
On-device execution that recomputes actions from live state and intent as conditions shift. It decides, trait by trait, how much of the expert's strategy to carry into a new setting.
Deterministic constraint enforcement. Clamps unsafe commands, monitors execution and writes audit logs.
Not a policy library. An execution loop driven by inferred intent, with hard constraint projection.
One engine. Three places we are pointing it.
Skilled pilots handle what autopilots find hardest: changing payloads, gusts, tight spaces and fast, unpredictable targets. That skill stays with the pilot. RAST is being developed to infer a pilot's strategy and carry it into autonomous flight across airframes, payloads and conditions, with every command bounded by the safety layer. First targets: carrying one pilot's strategy across airframes and payloads, close-proximity inspection in tight or GNSS-denied spaces, and terminal engagement against drones that evade.
Surgical robots record every motion of every expert. That makes them the best place to test whether an individual's motor strategy can be inferred, and split into what stays stable and what adapts to the task. We are working from recorded expert surgical motion first, because it is the cleanest place to prove the method. The specific clinical and platform focus will follow.
The same engine, pointed at music. Blizbeat reads how you move and turns it into a beat with your timing in it. No science claims here. It is a toy that happens to share the core code.
State-aware control, in the loop
- Objective inference from experts: infer what experts optimize for, not only what they did.
- Trait and state: separate what stays stable in an operator's strategy from what adapts to the situation. The stable part is what we transfer.
- State-aware action computation: actions are computed continuously from live state, not replayed from a script.
- Minimal-change corrections: adjust only what matters to the task and preserve the rest of the natural strategy.
- Hard safety limits: bounds enforced in the loop, with audit logs and predictable behavior.
We do not claim magic. We engineer the bridge from movement models to systems that run in the loop.
Team
I have spent decades building high-reliability systems where correctness and operational reality matter, as CTO of Oracle Israel's Enterprise Division and Chief Architect at Amdocs. Today I apply that mindset to Physical AI, turning movement science into software that runs in the loop, safely and measurably.
Contact
Building autonomous drones or surgical robots, and want expert skill to survive new conditions? Let's talk about data, pilots and integration.