Venue: CoRL 2025 (Award Finalist) · arXiv: 2506.15799 Authors: Andrew Wagenmaker, Mitsuhiko Nakamoto, Yunchu Zhang, Seohong Park, Waleed Yagoub, Anusha Nagabandi, Abhishek Gupta, Sergey Levine (UC Berkeley · University of Washington · Amazon) Category: RL for VLA / Diffusion Policies Trend tag: RL without log-probs

flowchart LR
  Z[Initial noise z0] --> FROZEN[Frozen diffusion policy]
  FROZEN --> A[Action chunk]
  A --> R[Environment reward]
  R --> LEARN[Train a policy over z0<br/>RL action space = noise latent]
  LEARN --> Z

Diffusion / flow-matching policies don't expose action log-probabilities, so the standard RL toolkit (policy gradient, PPO) can't be applied out of the box. Prior workarounds either finetune the whole model (unstable) or approximate log-probs (noisy).

DSRL (Diffusion Steering via Reinforcement Learning) makes a structural observation: the initial noise fed into a diffusion policy largely determines the sampled trajectory. Therefore you can freeze the diffusion policy and train a small RL actor that chooses the initial noise z₀ given an observation. The noise latent becomes the RL action space — with real log-probs and cheap rollouts. Concretely, DSRL trains a lightweight off-policy actor-critic (SAC) over the latent-noise input, requires only black-box access to the BC policy (it never touches the base weights or gradients), and adds only a small set of trainable parameters (~500K).

Post-hoc improvement of frozen diffusion policies across simulated benchmarks (OpenAI Gym, Robomimic, OGBench), single-task and multi-task (BridgeData V2) diffusion policies, and the pretrained generalist π₀ policy (steered from its public DROID weights). DSRL reports substantially higher sample efficiency than direct-finetuning RL baselines (e.g., RLPD), and demonstrates real-world autonomous improvement on a physical Franka setup. CoRL 2025 Award Finalist.

The seed idea for the ICLR 2026 wave of RL-for-flow-matching papers. Every approach on the ICLR 2026 RL topic page can trace an ancestor to DSRL or its contemporaries:

• π*0.6 + RECAP — advantage-conditioned (not log-prob-free via noise, but log-prob-free via positive/negative labels)
• RL Tokens (RLT) — compact RL-token interface over a frozen VLA
• SimpleVLA-RL — scalable open-source RL framework
• VLA-RFT — RL in a world model
• Stage-Aware RL, VITA

• arXiv: https://arxiv.org/abs/2506.15799

• Survey (CoRL 2025) · RL for VLA

← Back to CoRL-2025