Research map / Reinforcement learning / Visual and representation RL

Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry

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 Implicit Drifting Policy (IDP), a one-step imitation learning policy that trains with a geometry-aware quadratic potential whose weights come from observation-conditioned expert action variance compared against a global reference, plus an expert-proximal probe.

Problem

One-step generative action policies lose intermediate action correction, and explicit drifting fields are ill-posed in behavior cloning because each observation has only one expert action (conditional empirical distributions are Dirac).

Approach

Replace explicit drifting-field regression with an implicit quadratic correction potential whose diagonal weights are the excess precision of observation-similar expert actions relative to a global reference geometry (ReLU-truncated), and evaluate the generator near expert actions during training.

Key results

• On 3D pointcloud manipulation (56 tasks, 3 seeds), IDP averages 81.3% vs. Naive Drifting 79.8% and OFP 71.6% (both one-step).
• On 2D state-based tasks, IDP averages 0.87 (max) vs. Naive Drifting 0.81 and MIP 0.85 (2 NFE), matching multi-step DP at 0.87.
• Real-world Pick Peach (vision): IDP reaches 50% success while DP, Naive Drifting, and MIP all reach 0%.

Surprising findings

• In the real-world state-based (qpos) setting, IDP (30%) underperforms MIP (40%), which the authors attribute to joint-angle proximity being a poor proxy for task context.
• Removing the expert-proximal probe drops ToolHang-State success from 81.0%/60.0% to 40.0%/25.5%, below Naive Drifting.

Limitations

• Single-step policies struggle to cover multimodal action distributions; real-world bimodal task collapsed to one placement mode.
• Conditional expert geometry relies on a reasonable density of expert actions under similar observations.
• Diagonal approximation drops cross-coordinate coupling of the excess geometry.

Experimental setup

Datasets: Robomimic Tool-Hang PushT Adroit DexArt MetaWorld real-world Pick Peach (ALOHA)

Benchmarks: Robomimic Tool-Hang PushT Adroit DexArt MetaWorld Pick Peach real-world

Metrics: success rate max performance / average of last 5 checkpoints

Baselines: DP Flow DP3 CP MIP MP1 OneDP OFP FlowPolicy Naive Drifting

Builds on: Generative Modeling via Drifting MIP Diffusion Policy Consistency Policy MeanFlow

Closest papers by meaning

• FlowPolicy: Enabling Fast and Robust 3D Flow-Based Policy via Consistency Flow Matching for Robot Manipulation (AAAI 2025 · similarity 0.90)
• One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.89)
• Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.89)
• Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.89)
• Latent Diffusion Planning for Imitation Learning (ICML 2025 · similarity 0.89)
• Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.89)