Tsinghua's SAFLOW policy, accepted at NeurIPS, tops all 8 IsaacLab embodied control tasks
青稞AI · wechat · 2026-10-21
A Tsinghua University team led by PhD student Zhan Guojian published SAFLOW (Self-Adjoint Flow Policy Optimization), accepted at NeurIPS 2026 with OpenReview scores of 5/5/5/4.
The paper targets the likelihood blind spot of generative policies in online RL: flow-matching policies express rich multi-modal action distributions, but standard explicit Euler discretization is not time-symmetric, so forward sampling and reverse likelihood estimation diverge numerically, corrupting policy ratios. Inspired by symplectic integration, SAFLOW builds a self-adjoint policy flow via Verlet kick-drift-kick splitting, making action sampling and likelihood evaluation exactly invertible under the same discrete rules, with second-order consistency propagating to the optimization objective.
Results: highest average return on all 8 NVIDIA IsaacLab continuous control tasks (Ant, Humanoid, FrankaArm, ShadowHand, Unitree-Go2/H1, Anymal-D, Quadcopter), with Humanoid improved 50.9% over the best baseline. Advantages grow with GPU parallelism (128→4096 environments), and ablations confirm the self-adjoint structure drives the gains. The team positions SAFLOW as a continuous action head for VLA models and a basis for online post-training of embodied agents.
More from Embodied
- Sergey Levine: robotics is repeating LLMs' shift from fine-tuning to plain-language prompting — adityaag · 2026-10-23
- Embodied First Exam benchmarks coding agents on 1,090 tasks across 44 robots — ManlingLi_ · 2026-10-23
- Delivery robot spotted cruising city sidewalks on its own — HomeNowWTF · 2026-10-23
- Dev builds ESP32 desk display showing Claude Code and Codex agent status via local Go helper — tonylook · 2026-10-23
- Dream Labs emerges from stealth with Dream-1 robot foundation model at $3/hour — scott_e_reed · 2026-10-23
- AdaDex turns seconds of human hand motion into generalizable dexterous robot skills — mangahomanga · 2026-10-23