Same policy can fail outside its training engine: Manda tests sim-to-sim transfer across MuJoCo, PhysX, Newton and Genesis
facontidavide · x · 2026-11-03
Manda Robotics trained robot policies in four physics engines — MuJoCo, PhysX, Newton and Genesis — and tested how well they transfer across simulators.
Key finding: even with identical assets, the choice of physics engine changes learned behavior and success rates. A policy may simply fail once taken out of the engine it was trained in. The result suggests the simulator itself is an underrated variable in sim-to-real research: engine-specific dynamics get baked into the trained policy.
More from Embodied
- Hugging Face joins Asimov 1, building on the open robotics platform — mishig25 · 2026-11-03
- Scoble shares on-demand CNC printing with agents making parts overnight — Scobleizer · 2026-11-03
- China replicates robot box-folding demo 3 months later with much cheaper hardware — Scobleizer · 2026-11-03
- Unitree G1 learns self-rescue via ROOF RL strategy using geometric rewards, no maps — Scobleizer · 2026-11-03
- SO-101 robot arm learns ping pong, best rally hits 23 bounces — Scobleizer · 2026-11-03
- Investor predicts a $2-3k clothes-folding home robot within two years — lukas_m_ziegler · 2026-11-03