Discrete action representations match or beat continuous ones for robot policies
GuanyaShi · x · 2026-10-21
- Second paper in the "Robot Learning Science" series challenges the common belief that robot policies need continuous action heads.
- Key findings:
- Even the simplest discrete representation (512 bins per action dimension) matches or beats continuous actions across single-task and generalist policies, sim and real, single-step and diffusion/flow settings.
- Continuous regression/flow policies over-smooth action chunks: in RoboCasa, predicted jerk was only 7% of expert actions', while discrete methods stay much closer to expert dynamics.
- Conclusion: discrete actions are not less natural or less precise—they may be the better, more robust choice.
More from Embodied
- Plato ring launches at $199 to turn near-silent speech into AI commands — TinfoilTricorn · 2026-10-21
- Dream-1 controlled 31 unseen robot types across 53 countries within 24 hours — gharik · 2026-10-21
- Stanford's FACT model uses human force signals to beat T-Rex by up to 60 points on dexterous tasks — stanfordnlp · 2026-10-21
- Stanford's T2-z lets robots probe hidden physics with a poke, lifting success from 34% to 73% — stanfordnlp · 2026-10-21
- Snap Hires Robert Downey Jr. as Chief Customer Experience Officer to Sell Specs AR Glasses Against Meta — adariostrange · 2026-10-21
- Upcoming paper claims big models can handle latency-critical real-time robotics — ManlingLi_ · 2026-10-21