Neural Action Codec: compressing robot actions like audio to train VLA policies
YuXiang_IRVL · x · 2026-09-05
Researcher @ahadj0 introduces Neural Action Codec (NAC), a method inspired by audio waveforms that builds a compressed action vocabulary for training VLA (vision-language-action) policies. The work will be presented at CoRL in Austin, November 9-12, and the author is looking to connect with others working on scaling robot foundation models and multimodal tokenization.
More from Embodied
- Voice-summoned robot vacuum you can point at spots shows how fast embodied AI is landing — omooretweets · 2026-09-05
- Robot boxing match streams live telemetry via Foxglove; Agent auto-diagnoses the downed bot — jonstephens85 · 2026-09-05
- Dex4D Combines Point-Track Policy and Video-Gen Planners for Sim-to-Real Dexterous Hands — chris_j_paxton · 2026-09-05
- Cybercab goes live in Austin as Tesla hires AI engineers across the board — elonmusk · 2026-09-05
- Musk shares Cybercab measurement: just ~40 dB at a cruising 40 MPH — elonmusk · 2026-09-05
- Reachy Mini robot skit ep.2: breaking bad news to Microduck with new emotion approach — Thom_Wolf · 2026-09-05