🤓 Latest Submissions
JACOBI
JACOBI teaches two simulated SO-101 robot arms to clear a dinner table from natural language and four camera views. The task is bimanual and contact-rich: open the tabletop drawer, retrieve spoons and forks, pick up two plates and a mug, and place every object in its target zone — upright, separated and supported by the table. Twelve joints are commanded; four synchronized 320x320 camera views and a 12-dimensional joint state are observed. The project's central claim is that the grasping is genuinely physical. The jaws are position servos that drive into an object and stall against it; the resulting contact and friction forces are the entire grip. There is no weld constraint, no equality attachment, no object-follow and no pose teleport anywhere in the transport path. Objects carry real mass, inertia and gravity. If friction is insufficient, the object falls and the episode is rejected. Plates are the hardest case, so their collision shape is a ring of capsule segments around the rim rather than a solid disc — that is what lets parallel jaws pinch them reliably. Training data was generated in MuJoCo by a scripted expert, and every episode was accepted only after passing physical success checks: measured object lift, drawer travel, tabletop support and zone containment. The merged dataset holds 71 successful episodes and 218,384 frames at 25 Hz in LeRobot v3 format. We are fine-tuning SmolVLA on this dataset, adapted from the base model's three-camera layout to our four. A 20-step smoke test across all 71 episodes brought loss from approximately 4.5 to 1.13. The deployment phase targets an Intel Core Ultra 7 265T. We are exporting the policy for OpenVINO, benchmarking FP32, BF16, FP16 and validated INT8 variants across CPU, integrated GPU and NPU, and replaying held-out tasks to confirm that optimization has not damaged task success. That evaluation is in progress; we are not yet quoting latency or success figures.
16 Sep 2026
