Research map / Reinforcement learning / Visual and representation RL

Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control

NeurIPS 2026 · Reinforcement learning › Visual and representation RL · approach: Skill learning and diffusion policies · 0 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.

Proposes GeCO, a time-unconditional flow-matching policy head that learns a stationary action-sequence velocity field, used for gradient-based adaptive-step inference and field-norm OOD detection, evaluated on LIBERO, RoboTwin 2.0, VLABench and real-world Galaxea tasks.

Problem

Time-conditioned diffusion and flow-matching policies use a fixed integration schedule regardless of state difficulty and lack an intrinsic energy-like signal for action validity or OOD detection.

Approach

Drop time conditioning and learn a single stationary velocity field over action sequences with a truncated-linear rescaling c(gamma) that vanishes at expert actions; inference runs gradient-style updates until the field norm falls below a threshold, and the final field norm serves as a training-free OOD score.

Key results

• On LIBERO, GeCO with 5 max steps reaches 91.9% average success vs 90.0% for rectified flow at 20 steps, using 5.0 NFE; at 20 steps it reaches 93.5%.
• On RoboTwin 2.0 Hard, pi0 with GeCO raises average success from 0.18 to 0.28; on real-world Nut Assembly success rises from 30% to 70% with NFE 10.0 to 5.06.
• OOD detection on LIBERO Goal-to-Spatial reaches AUROC 0.93 vs 0.53 for the flow-matching loss baseline.

Surprising findings

• Fixed constant step-size schedules (e.g., 0.02 x 20) perform nearly as well as the adaptive schedule (93.05 vs 93.5 average), suggesting gains come mainly from the stationary field.
• On LIBERO, the GeCO pi0 variant is slightly below the standard pi0 baseline (93.9 vs 94.2), and it is also below pi0.5 in the comparable setting.

Limitations

• Theoretical foundations (convergence, bounds) are acknowledged as underdeveloped.
• OOD evaluation is limited to a single suite-level shift (LIBERO Goal to Spatial).
• Gains on several LIBERO suites over the flow-matching baseline are absent; GeCO is comparable or slightly lower on the pi0 LIBERO average.

Experimental setup

Datasets: LIBERO RoboTwin 2.0 VLABench teleoperated real-world demonstrations

Benchmarks: LIBERO RoboTwin 2.0 VLABench real-world Galaxea R1 Lite tasks

Metrics: success rate NFE inference time AUROC TNR TPR time saved

Baselines: Diffusion Policy Rectified flow pi0 pi0 + FAST pi0.5 ACT G0 Plus Diff-DAgger-style flow-matching loss

Builds on: flow matching rectified flow pi0 CleanDiffuser Diff-DAgger

Closest papers by meaning

• Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.90)
• ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
• Aligning Flow Map Policies with Optimal $Q$-Guidance (NeurIPS 2026 · similarity 0.89)
• FreqPolicy: Efficient Flow-based Visuomotor Policy via Frequency Consistency (NeurIPS 2025 · similarity 0.89)
• Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.89)
• Conditioning Matters: Training Diffusion Policies is Faster Than You Think (NeurIPS 2025 · similarity 0.88)