Mumbai Coder

LookAct is an embodied AI agent that checks its own work before claiming success. Vision-language agents that act in the physical world usually look once, act once, and report “done”, so when the target is ambiguous, the scene changes, or a motion fails, they are confidently wrong. LookAct runs a verification loop called TrustLoop: observe, interpret, plan, validate, act, verify, recover. It asks the user instead of guessing between similar objects, never acts on a camera frame taken before its last movement, detects when a target has moved and replans, and only reports success after a fresh verification frame confirms it. A deterministic safety layer validates every motion command, and the lidar can stop the robot at any time. The AI never controls the motors directly. It runs on a TortoiseBot (Raspberry Pi 4, ROS 1 Noetic, lidar, Pi camera, with an OAK-D depth camera planned), with a live web dashboard and a built-in simulator that compares TrustLoop against a conventional one-shot agent using the same models. This is a work in progress: the agent, safety layer, simulator, and camera pipeline on the robot are running, and full real-robot driving trials are next.