.png&w=256&q=75)
🤓 Latest Submissions
Bimanual VLA with OpenVINO on Core Ultra
Intel Bimanual VLA is a language-driven bimanual manipulation system built in seven days by one developer new to robotics. It targets the Intel Core Ultra 7 358H and deploys through OpenVINO on CPU, iGPU and NPU. Four skills run end to end: pick(fork), place(fork), pick(water_bottle) and handoff(A to B, fork). A spoken command reaches execution through the full path. Speechmatics transcribes it, a rule-based grammar with 43 tests grounds it, and a scripted controller runs it. run_demo.py completes 4 of 4 skills in about 25 seconds from a clean clone. Intel hardware carries the perception stack. PoseNet trains to 2.6-3.2 mm MAE on the Arc B390 iGPU. OpenVINO then runs it at 1465 Hz on GPU FP16, 910 Hz on NPU and 5800 Hz at batch 16, across three devices and three precisions. Grasping is an explicit MuJoCo weld constraint rather than friction contact, disclosed as a modelling choice (ADR-029, ADR-015). Every claim is measured, and reported with the qualifier that makes it true. Across 20 seeds per skill, pick(fork) and place(fork) reach 20/20. pick(water_bottle) reaches 9/20. An earlier 10-seed run gave 6/10, and we report the larger sample instead. handoff also reaches 20/20, and we label that figure degenerate: its envelope is a single point, so the number measures determinism rather than robustness. Three findings changed a decision. INT8 quantization was declined, because it deviated 36-37 mm against the model's own 2.6-3.2 mm MAE. A geometry redesign was rejected on its own evidence: two of three kinematic walls dissolved, but only 1 of 8 skills passed. A motion-stack rewrite then produced a six-stage bimanual relay that passes all five scene-integrity criteria at every stage, moving the fork between arms via the table. Direct hand-to-hand transfer remains open, diagnosed, with the fix named in ADR-073. 59 ADRs record every decision and every reversal. SUBMISSION.md carries the full scope, including what does not work.
16 Sep 2026
