Fine-tuning VLAs on TAMP data hits distribution mismatch; DATAFARM gets JAX port
tomssilver · x · 2026-10-28
Thomas Silver notes that fine-tuning VLAs on TAMP-generated data isn't the easy win it appears to be: the key issue is distribution mismatch with pretraining data. Samrat Sahoo's clever workaround learns to align TAMP data to the pretraining distribution, and it works well.
A lightweight JAX/PyRoki port of DATAFARM (DataFarm-PyRoki, MIT license) is also available, adding the frozen motion encoder with pretrained weights, a whole-motion-leg residual with constrained trajectory solving, trajectory retiming, and joint-density-based IK selection — making procedurally generated, humanlike robot motion data easier to use outside the TiPToP framework.
More from Embodied
- Testing Mint Modeler to Build the Menlo Research Robot — freelerobot · 2026-10-28
- Khosla-backed startup builds a 'mind-reading' cap with 100,000 tiny chips to write AI prompts — Distinct-Question-16 · 2026-10-28
- Racer turns Ray-Ban Meta glasses into an AI track coach with Opus 5.5, gains 5.3 seconds per lap — Vjeux · 2026-10-28
- Google brings Gemini Robotics ER 2 embodied reasoning model to Google Cloud — ZoubinGhahrama1 · 2026-10-28
- AI-Written Firmware Sparks Open-Source Music Hardware Wave as Behringer Goes Open — id-ltd · 2026-10-28
- StackChan robot adds expressive faces per context with Hermes-Agent, open source coming — alexcovo_eth · 2026-10-28