Intel — Bimanual VLA Manipulation. Plus Speechmatics for voice. We run a language-conditioned bimanual manipulation policy on a $400 Intel laptop with no discrete GPU — 12th Gen Core i3-1215U, UHD integrated graphics — and make it fast enough to close a robot control loop. Measured baseline on that machine: a 52M-parameter ACT forward pass takes 1,050 ms. We export to OpenVINO IR, compress weights to INT8 with NNCF, and run across CPU and the integrated GPU. Every configuration is scored on two axes over identical seeds: inference latency and closed-loop success rate on AlohaTransferCube-v0, a bimanual hand-off one arm cannot solve alone. A quantization that halves latency and drops success is reported as a failure, not a win. Speechmatics turns speech into the instruction the policy is conditioned on, so what you say changes what the arms do. Solo team, already building. Looking for: a robotics/VLA person, and a front-end dev for the demo UI.
.png&w=256&q=75)
Founder