AI Infra Summit Hackathon · Intel Bimanual VLA Manipulation

Bimanual VLA on Intel

A measurement, not a demo.

70%
ties ACT ceiling, 20 fixed seeds
90.5 ms
INT8 on Intel iGPU (6x vs PyTorch CPU)
0/10
swap-test effect — language never steers behavior

Training curve: 20% → 70% over 200k steps

CheckpointSuccess (20 seeds)
15,000 steps4/20 (20%)
125,000 steps9/20 (45%)
150,000 steps13/20 (65%)
175,000 steps13/20 (65%)
200,000 steps14/20 (70%)
ACT baseline (no language)14/20 (70%) — ceiling

The instruction-swap test: does language actually steer behavior?

Feed the model the wrong instruction while running the same task. If language matters, success should collapse.

CheckpointNormal instructionSwapped instructionVerdict
125k steps9/20 (45%)12/20 (60%)WRONG DIRECTION
150k steps13/20 (65%)13/20 (65%)IDENTICAL SEEDS
200k steps14/20 (70%)14/20 (70%)IDENTICAL SEEDS
At every checkpoint tested, changing the sentence never changed a single episode's outcome — even at 200k, where the model ties ACT. Root cause: in the training dataset, every episode pairs one task with one instruction 1:1, so the visual scene alone fully determines behavior. Nothing in training ever rewards reading the sentence — this is a dataset confound, not a training-length problem.

Intel optimization: official benchmark_app numbers

RuntimeMedian latencyThroughput
fp32, CPU239.1 ms3.92 FPS
INT8, CPU195.9 ms4.20 FPS
INT8, Intel iGPU71.0 ms13.91 FPS

INT8 weights: 132.6 MB → 33.7 MB (3.9x smaller). Reproducible with benchmark_app -m ir/langact_int8.xml -hint latency -d GPU.
Hardware disclosure: 12th Gen Core i3 (Alder Lake), CPU + Intel UHD iGPU only — no NPU.

Voice control: since in-model conditioning is inert, we route instead

Speechmatics STT transcribes the command, keyword matching resolves it to a known task, and that task actually runs — Speechmatics TTS confirms the true outcome. Asking for a task the system can't do (peg insertion, measured 0/20) gets an honest decline, not a silent failure.

GitHub Repository → Trained Checkpoint (Hugging Face) → Slide Deck (PDF) →