Legged-robot state estimator from proprioception alone, leveraging contact events
chris_j_paxton · x · 2026-11-20
A collaboration with Oxford, SNU and Pisa released a legged-robot state estimator that uses proprioception only, carefully choosing when and how often to use contact events. Kyle Morgenstein notes many labs quietly hurt policy performance by reducing observation spaces without thinking about state estimation, and recommends trying this method. Code is open-sourced on GitHub.
More from Embodied
- Teamily launches Video Agent letting AI agents see through your AI glasses — CurieuxExplorer · 2026-11-20
- Verne Robotics' 8.2B open-weight V0.5 hits 99.3% on LIBERO, dubbed the Thinking Machines of robotics — chris_j_paxton · 2026-11-20
- 60-Day Physical AI Challenge: Simulating a Robot Arm in MuJoCo, No Hardware — DynamicWebPaige · 2026-11-20
- Autonomous Cable Picking Live Demo on ABB Cobots, Dec 2 — stepjamUK · 2026-11-19
- Musk details Terafab: 100M sq ft fab targeting first test chip by June 2027 — FuSheng_0306 · 2026-11-19
- Musk: Grok 4.8 pretraining done, Optimus to show unscripted capabilities within 1-2 months — FuSheng_0306 · 2026-11-19