Event
OrbiSim proposes a differentiable physics engine simulation paradigm
1 reports · 1 sources
Jiajian Li et al. propose OrbiSim, a robot simulation paradigm that reconstructs the world model as a fully differentiable physics engine, connecting structured scene assets, neural dynamics, and downstream reinforcement learning. The entire simulation loop, from explicit state to visual observation generation, supports end-to-end differentiability, and can be used for tasks such as differentiable contact modeling, gradient-based policy optimization under sparse rewards, and physical reasoning. The paper claims that experiments show OrbiSim significantly outperforms existing world models in prediction fidelity and control performance, and responds consistently to asset configurations and physical parameters. The paper was released on arXiv cs.RO, with the latest version being v2 dated October 16, 2026.
Report timeline
- OrbiSim: Recasting World Models as a Differentiable Physics Engine for Embodied Intelligence
OrbiSim proposes a robot simulation paradigm that recasts world models as a fully differentiable physics engine, connecting structured scene assets, neural dynamics, and downstream reinforcement learning. The entire simulation loop, from explicit state transitions to visual observation generation, supports end-to-end differentiability, enabling tasks such as differentiable contact modeling, gradient-based policy optimization under sparse rewards, and physical reasoning. Experiments show that OrbiSim significantly outperforms existing world models in prediction fidelity and control performance, and responds consistently to asset configurations and physical parameters.
arXiv cs.RO ·