🤓 Latest Submissions
Command - NL Control Layer for Legged Robots
Legged robots already have working motion primitives - gait controllers, RL policies, canned animations. What they don't have is a universal natural-language front end that reasons about whether a command is actually safe to run right now, given the robot's live state. Command is that layer. The same literal command, "walk forward 2 meters," produces three different outcomes depending on state: at 90% battery on flat terrain it executes in full; at 30% battery on rough terrain it's modified to a shorter, safer distance with a stated reason; at 8% battery it's refused outright, citing the failed precondition. This isn't keyword matching - we proved it live against commands a regex baseline structurally cannot solve: the model refuses "move forward until you find something weird" because the robot's capability profile has no perception primitive, plans an 8-step square from "walk a square, 1 meter each side," and computes verifiable battery-budget arithmetic for round-trip commands (confirmed by post-run telemetry matching its own math). A deterministic validation layer sits between the model and the robot: hallucinated function names, bad parameters, and preconditions that fail mid-plan are all caught before any primitive reaches the hardware. The same reasoning layer dispatches, unchanged, across a hexapod, a quadruped, and a wheeled rover with an entirely novel capability profile - proving the interface is genuinely robot-agnostic, not hardcoded per morphology. Model inference runs on AMD Instinct GPUs via the AMD Developer Cloud (ROCm), with Fireworks-on-AMD as the AMD-powered development backend - switching between them is a config change, not a code change. A live PyBullet simulation and a real physical hexapod (driven over its existing WiFi protocol, no firmware changes) both demonstrate the same validated reasoning output in motion.
13 Jul 2026
