TaskForge VLA

TaskForge VLA is a simulation-first embodied AI system built for the AI Infra Summit Hackathon 2026, targeting the Intel Bimanual VLA Manipulation with Multi-Modal Reasoning track. Users can give natural-language commands such as “Pick both blocks” or “Put the red block near the blue one.” TaskForge interprets the instruction with an OpenRouter-powered LLM, converts it into a typed, structured goal, and falls back to a deterministic parser if needed. The system observes the simulated world, generates a symbolic plan, and uses a reachability-aware bimanual coordinator to decide which arm should perform each action. Before execution, a deterministic validator checks objects, actors, dependencies, grasp states, and supported actions. Validated plans are executed in MuJoCo using deterministic primitives such as pick, place, and move-home, with analytical inverse kinematics controlling the arms. After execution, TaskForge re-observes the world and verifies whether the original goal is satisfied using predicates such as `left_of`, `right_of`, `near`, and `grasped_by`. A key feature is closed-loop recovery. If an action fails, TaskForge preserves completed work, observes the updated world, creates a residual goal for only the unfinished task, replans, validates the new plan, and continues execution. The project also includes a Streamlit dashboard for task input, plans, world state, execution status, recovery events, and final verification, plus a live MuJoCo viewer for visual robot simulation. Core principle: AI interprets. Structured systems validate. Deterministic robotics executes. The environment decides success.