InfraForge VLA: Language to Robot Action

InfraForge VLA is a language-controlled robotic manipulation pipeline built for Intel's Bimanual VLA Manipulation Online Challenge, part of the AI Infra Summit Hackathon. How it works: a plain-language instruction such as pick up the mug with arm A is parsed by a rule-based language interface into a structured action plan. That plan drives a simulated bimanual robot, built on MuJoCo and LeRobot's ALOHA environment, through a real pick-and-lift sequence using end-effector control. What is verified, not just claimed: pick-up succeeds and is reproducible, reaching reward 2 out of 4 for the right arm and 4 out of 4 for the left arm, both confirmed on real hardware. The pick-up mechanic is position-relative, computed from the object actual position using a measured geometric offset rather than hardcoded coordinates. We validated that this generalizes with a 10-seed randomized placement test, succeeding on 9 out of 10 seeds within an empirically measured safe range. Inference is benchmarked with Intel OpenVINO on real CPU and integrated GPU hardware, including an honest finding that for this small policy model, CPU outperforms GPU, since GPU dispatch overhead dominates at this model size. What is honestly still in progress: two-arm hand-off, passing the object between grippers, is not yet reliable. We root-caused the actual blocker through direct physics measurement rather than guesswork, documented in full in the repository technical notes. We prioritized a smaller, genuinely working, well-documented pipeline over a larger scenario attempted unreliably, given the online-only timeline.