CoAct — Consensus-Gated Bimanual Manipulation

CoAct is a consensus-gated bimanual manipulation system built around a simple question: what should a robot do when two embodied viewpoints produce different interpretations of the same human instruction? In shared environments, spatial language can be ambiguous. Two robotic arms observing the same workspace from different viewpoints may ground an instruction to different objects or locations. Instead of immediately acting on one interpretation, CoAct treats this disagreement as a safety signal. Each SO-101 agent independently observes and grounds the instruction. A consensus engine compares their interpretations before physical execution is permitted. When disagreement occurs, CoAct diagnoses the ambiguity and attempts to resolve it through re-observation and confirmation. Its key interaction mechanism is live-correctable hover confirmation. Before committing to manipulation, the robot approaches the interpreted target while remaining interruptible. Speechmatics real-time transcription listens for human corrections during this window. If the user says, “No, the large plate,” CoAct can detect the correction, abort the current motion, retract, update the intended reference, re-ground the instruction, and evaluate consensus again before continuing. The prototype combines bimanual SO-101 robot models, MuJoCo simulation, independent grounding, consensus reasoning, ambiguity diagnosis, inverse-kinematics motion control, a safety-oriented state machine, asynchronous event handling, and Speechmatics real-time ASR. CoAct introduces a deliberate checkpoint between interpretation and physical commitment, enabling robots to recognize conflicting interpretations and remain responsive to human correction. **CoAct: Consensus before commitment. Correction before consequence.**