Research map / Reinforcement learning / Visual and representation RL

Conditioning Matters: Training Diffusion Policies is Faster Than You Think

NeurIPS 2025 · Reinforcement learning › Visual and representation RL · approach: Skill learning and diffusion policies · 1 citations (OpenAlex)

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.

Identifies "loss collapse" in conditional flow-matching diffusion policies and proposes Cocos, which makes the source distribution condition-dependent (Gaussian centered on an autoencoded condition embedding), achieving faster training and higher success on LIBERO, MetaWorld, and real robots.

Problem

Conditional diffusion/flow-matching policies for VLAs train inefficiently because, when conditions are hard to distinguish, the objective degenerates into modeling the marginal action distribution (loss collapse), so the policy ignores conditioning inputs.

Approach

Replace the standard Gaussian source distribution in conditional flow matching with a condition-dependent Gaussian N(alpha*F(E(c)), beta^2 I), where F is an autoencoder-trained encoder on frozen DINOv2/T5 condition embeddings, so the source is anchored to condition semantics and the policy must use the condition.

Key results

• On LIBERO, Cocos (0.2 std) averages 94.8% success vs 86.5% for DP-DINOv2 and 64.4% for DP-scratch.
• Cocos reaches high success within 5-10K steps, reported as roughly 2.14x faster convergence than DP-DINOv2; the paper says it matches pi0 and OpenVLA-OFT within about 30K gradient steps.
• On MetaWorld (30 tasks), Cocos averages 74.8% vs 59.5% without Cocos.

Measured effects reported in the paper

Surprising findings

• Cocos with beta=0.1 (more concentrated source) performs much worse on LIBERO (77.5% average) than beta=0.2 or 0.4, indicating overly concentrated condition-dependent sources hinder training.
• Policies without Cocos tend to ignore 3rd-person camera views, relying almost only on the wrist camera.

Limitations

• The source distribution is instantiated only as a fixed-variance Gaussian from autoencoding conditions; more expressive priors are unexplored.
• Experiments focus on multi-task imitation learning; large-scale pretraining and broad VLA generalization are not evaluated.
• Real-world evaluation uses only 10 trials per task on two robot platforms.

Experimental setup

Datasets: LIBERO MetaWorld

Benchmarks: LIBERO MetaWorld SO100 real-world tasks xArm real-world tasks

Metrics: success rate cosine similarity of hidden states norm scale change

Baselines: DP-scratch DP-DINOv2 pi0 OpenVLA-OFT Seer

Builds on: conditional flow matching RDT pi0 DINOv2 T5

Closest papers by meaning

• Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.90)
• Diffusion Policy Policy Optimization (ICLR 2025 · similarity 0.89)
• Text-Aware Diffusion for Policy Learning (NeurIPS 2024 · similarity 0.89)
• DiffusionVLA: Scaling Robot Foundation Models via Unified Diffusion and Autoregression (ICML 2025 · similarity 0.89)
• Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.88)
• One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.88)