Research map / Reinforcement learning / Imitation learning

SkiP: When to Skip and When to Refine for Efficient Robot Manipulation

NeurIPS 2026 · Reinforcement learning › Imitation learning · approach: Behavior cloning theory and models · 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 SkiP, which relabels behavior-cloning targets so skip-segment steps predict the next key-segment entrance, with MSK (DCT-based motion spectrum keying) producing key/skip labels; reduces executed steps 15-40% while matching or improving success across RLBench, RoboMimic, RoboTwin and real-robot pi0.5 fine-tuning.

Problem

Imitation learning policies predict and execute actions uniformly at every control step, which wastes decisions on redundant free-space motion and compounds prediction error, while contact-rich phases need dense refinement.

Approach

Action relabeling: in skip segments the behavior-cloning target is shifted to the action at the entrance of the next key segment, while key segments keep next-step targets; key segments come from a task-agnostic DCT spectral procedure (MSK) plus bend and keyframe heuristics, all within one unified network.

Key results

• On RLBench-10, SkiP reaches 0.850 average SR with 72.9 steps, +0.149 SR over CoA-rev and 54.6 fewer steps.
• On RLBench-18 fine-tuning from pi0.5, SkiP improves average SR from 18.59% to 20.96% and reduces Steps_succ from 108.30 to 66.38.
• On RoboMimic, SkiP reaches 0.772 average SR versus 0.723 for CoA-rev and cuts Steps_succ by 32% relative to CoA-rev.

Surprising findings

• Skipping improves accuracy: collapsing free-space traverses into one decision removes intermediate prediction errors, so SR often rises while steps fall.
• VLM-derived key segments (Gemini-2.5-Pro) perform much worse than MSK on RLBench and RoboMimic, suggesting semantic phase boundaries are not precise enough.

Limitations

• Requires absolute action targets for relabeling.
• Performance depends on the label source; VLM or heuristic segmentation can underperform.
• Some tasks remain at 0% SR for both methods, e.g. on RLBench-18 fine-tuning.

Experimental setup

Datasets: RLBench RoboMimic RoboTwin 2.0 Open X-Embodiment (pi0.5 pretraining)

Benchmarks: RLBench-10 RLBench-60 RLBench-18 RoboMimic RoboTwin 2.0 real-robot tabletop tasks

Metrics: success rate Steps Steps_succ wall-clock time

Baselines: DP ACT CoA-fwd CoA-rev KF-only DP3 Base pi0.5

Builds on: Chain-of-Action Diffusion Policy DP3 pi0.5 ACT

Closest papers by meaning

• Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.89)
• Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.88)
• Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.88)
• Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation (NeurIPS 2025 · similarity 0.88)
• Data Scaling Laws in Imitation Learning for Robotic Manipulation (ICLR 2025 · similarity 0.88)
• Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling (ICLR 2025 · similarity 0.87)