Menlo Achieves Zero-Shot Sim2real for Asimov Biped, Citing Unitree Early-2025 Parity
chris_j_paxton · x · 2026-08-27
Menlo Research spent 8 months closing the locomotion sim2real gap on its bipedal Asimov robot and now claims zero-shot sim2real: policies trained in simulation run on real hardware without extra tuning.
Key points:
- Flat-terrain walking generalizes to slight slopes and grass textures, and works across multiple robot units, not just one hand-tuned lab unit
- Base policies run onboard at 50Hz on the robot's motion control board, driving 25 motors and sensors with no external compute
- The team argues the fix was less about model nudging and more about making the underlying hardware more deterministic
- They assess parity with Unitree locomotion circa early 2025, with fast progress remaining
- Learnings should transfer to other open-hardware robots
More from Embodied
- Indie Dev Rushing to Hardware Production: Sourcing Parts for First 50 Units — pramodk73 · 2026-08-27
- Skepticism on Video Prompts and Scaling Laws in Robotics — AvivTamar1 · 2026-08-27
- Mini-Pi Robot Attempts Long Jump: Tiny Machine Doing Its Best — TechPreacher · 2026-08-27
- Lab Uses Motion-Capture Suits to Teach Robots 'Positions' — tedmitew · 2026-08-27
- Gates should invest in farming robots instead of AI fear — mitchdeg · 2026-08-27
- Dev writes custom driver to boost global shutter FPS — yacineMTB · 2026-08-27