DATAFARM: aligning TAMP demos to VLA pretraining distribution unlocks planning plus dexterity
tomssilver · x · 2026-10-02
Task and Motion Planning (TAMP) excels at multi-step, semantic, and geometric reasoning, while VLAs excel at deformable object manipulation that TAMP can't do. DATAFARM tries to give one VLA both capabilities.
Tom Silver notes that fine-tuning VLAs on TAMP-generated data seems like an easy win but isn't—the key issue is distribution mismatch. Samrat Sahoo's workaround: learn to align TAMP data to the VLA's pretraining distribution before fine-tuning, which works really well.
More from Embodied
- City Farmers joins OpenAI's first Codex Physical Builds batch with a Raspberry Pi setup — broodsugar · 2026-10-02
- Five cars on one Austin block, four of them self-driving — binarybits · 2026-10-02
- Tesla adds 32 Cybercabs in Texas as its robotaxi rollout finally scales up — binarybits · 2026-10-02
- EMPIRIC teaches robots missing physics as code, solving all 25 tasks where baselines get 14-16 — tomssilver · 2026-10-02
- Agility's Digit runs end-to-end autonomous whole-body manipulation for ~12 hours at IROS — chris_j_paxton · 2026-10-02
- Tesla FSD slams the brakes just in time — driver in the next lane never noticed — xiaosun86 · 2026-10-02