Research map / Reinforcement learning / Policy gradient and actor-critic
Stabilizing Reinforcement Learning in Differentiable Multiphysics Simulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SAPO, a maximum-entropy first-order model-based actor-critic that uses differentiable-simulation gradients, and Rewarped, a parallel differentiable multiphysics platform for rigid bodies and deformables.
Problem
RL on tasks with deformable objects is limited by slow soft-body simulation and high sample complexity; existing first-order differentiable-simulation policy gradient methods get stuck in local optima due to non-smooth landscapes and contacts.
Approach
Adds entropy regularization to first-order analytic policy gradients from differentiable simulation (soft value targets, entropy-augmented returns, automatic temperature tuning) within an on-policy actor-critic, and implements a GPU-parallel differentiable MPM/FEM/rigid simulator in NVIDIA Warp.
Key results
- SAPO reaches 90.0 return on HandFlip vs 32.7 for SHAC and 38.2 for APG.
- SAPO achieves 4535.9 on AntRun vs 3621.0 for SHAC, and 1820.5 on SoftJumper vs 853.3 for SHAC.
- SAPO outperforms all baselines on all six Rewarped tasks at equal environment-step budgets, with 10 seeds each.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SAPO instead of SHAC | FO-MBRL on AntRun · AntRun | episode return: 3621.0 $\pm$ 54.4 -> 4535.9 $\pm$ 24.5 | Helped |
| Use SAPO instead of SAC | RL on HandReorient · HandReorient | episode return: 70.5 $\pm$ 10.2 -> 221.7 $\pm$ 9.5 | Helped |
| Use SAPO instead of APG | FO-MBRL on RollingFlat · RollingFlat | episode return: 86.9 $\pm$ 0.4 -> 100.4 $\pm$ 0.4 | Helped |
| Use SAPO instead of APG | FO-MBRL on SoftJumper · SoftJumper | episode return: 956.6 $\pm$ 15.6 -> 1820.5 $\pm$ 47.9 | Helped |
| Use SAPO instead of APG | FO-MBRL on HandFlip · HandFlip | episode return: 38.2 $\pm$ 3.5 -> 90.0 $\pm$ 2.2 | Helped |
| Use SAPO instead of SAC | RL on FluidMove · FluidMove | episode return: 28.2 $\pm$ 0.7 -> 30.6 $\pm$ 0.4 | Helped |
| Add entropy-augmented returns to actor training (ablation) | SAPO on HandFlip · HandFlip | episode return | Helped |
| Add design choices III, IV, and V: state-dependent policy variance, critic ensemble without target networks, and architecture/optimization changes (ablation) | SHAC on HandFlip · HandFlip | episode return | Helped |
| Add design choices III, IV, and V: state-dependent policy variance, critic ensemble without target networks, and architecture/optimization changes (ablation) | SHAC on DFlex rigid-body locomotion tasks · DFlex rigid-body locomotion tasks | performance | No clear effect |
Surprising findings
- APG outperforms SHAC on deformable tasks, while SHAC beats APG on articulated rigid-body tasks.
- Design changes applied to SHAC (state-dependent variance, no target networks, SiLU/AdamW) had minimal impact on DFlex rigid-body locomotion tasks.
Limitations
- Relies on subsampled non-occluded particle states as observations, infeasible for real-world settings.
- HandReorient only learns to catch the cube, not reorient it, within the budget.
- Differentiable rendering not yet combined with simulation, so no RGB/depth observations.
Experimental setup
Benchmarks: AntRun HandReorient RollingFlat SoftJumper HandFlip FluidMove
Metrics: episode return training stability across seeds
Baselines: PPO SAC TrajOpt APG SHAC
Builds on: SHAC Analytic Policy Gradient Soft Actor-Critic NVIDIA Warp DFlex
Closest papers by meaning
- Accelerating Visual-Policy Learning through Parallel Differentiable Simulation (NeurIPS 2025 · similarity 0.87)
- A Differential and Pointwise Control Approach to Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Staggered Environment Resets Improve Massively Parallel On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Rapidly Adapting Policies to the Real-World via Simulation-Guided Fine-Tuning (ICLR 2025 · similarity 0.86)
- Domain Randomization via Entropy Maximization (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.