Roboticist: needing ~10 future command frames for joint tracking signals a flawed formalism
KyleMorgenstein · x · 2026-11-20
Roboticist Kyle Morgenstein raises two objections in a debate with Chen Tessler and Chris Paxton over joint-space control:
- Joint commands under-specify intent: a joint command doesn't encode sufficient information about what we actually want the robot to do
- Future-frame dependency is a smell: many works show the smoothest joint tracking on hardware requires feeding the policy 10 future command frames — he argues a good formalism shouldn't need that, signaling the tracking-based approach is poorly formalized
More from Embodied
- Roboticist argues full joint space is the wrong abstraction for high-level agents — KyleMorgenstein · 2026-11-20
- Kyle Morgenstein: full joint space is the wrong abstraction for high-level robot agents — KyleMorgenstein · 2026-11-20
- US Physical AI Runs on Private CAPEX, Frontier Models, and Energy Flexibility — ingliguori · 2026-11-20
- Skild AI open-sources code, ending speculation about its robotics work at CoRL — KyleMorgenstein · 2026-11-20
- Skild AI releases Steel 0.2 robot simulation platform with 81 environments and magnetic contacts — KyleMorgenstein · 2026-11-20
- STASIS (YC F26) launches smart glasses with 6 sensors tracking eye strain at 92% accuracy — LinusEkenstam · 2026-11-20