Policy-DRIFT lands NeurIPS acceptance with 49% drag reduction, beating DRL by 16%
ricardovinuesa · x · 2026-09-28
Ricardo Vinuesa's team had Policy-DRIFT: Dynamic Reward-Informed Flow Trajectory Steering accepted at NeurIPS, targeting skin-friction drag in wall-bounded turbulence.
- Key idea: move reward information out of policy gradients and into generative inference — a conditional flow matching model builds a manifold of feasible flow states across control regimes, Terminal Reward Guidance steers samples toward reward-maximizing targets, and a lightweight DRL policy tracks the full-field targets via RMSE.
- Results: on the canonical Reτ=180 turbulent channel flow DNS benchmark, it achieves 49% drag reduction (near the theoretical upper bound), roughly 16% better than the DRL baseline, while using 37x less actuation energy.
More from Research
- KPI: a promptable interaction kernel lets humanoids do contact-rich tasks zero-shot — chris_j_paxton · 2026-09-28
- Self-Adaptive VLA: robots that learn from failed rollouts and fix themselves — gan_chuang · 2026-09-28
- Simon Prince Publishes Neural ODEs Tutorial: Residual Nets as Infinite-Layer ODEs — SimonPrinceAI · 2026-09-28
- Xiaomi open-sources 7,000+ RL environments with an interactive HF Space to run rollouts — SergioPaniego · 2026-09-28
- 77.8% of Agent Runs Carry Stale File Views — Lessons From Building a Context-Refresh Layer — jabulari · 2026-09-28
- GPT Researcher drops embeddings: LLM-based retrieval lifts relevant context 59% at same cost — hwchase17 · 2026-09-28