Research map / Image generation / Fast diffusion sampling
Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking
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 Self-Guidance (negative guidance from past-state conditioned predictions) and Adaptive Chunking (cosine-similarity-gated replanning) to improve consistency and reactivity of diffusion policies for generative behavior cloning, without extra models.
Problem
Diffusion policies for behavior cloning suffer from low-fidelity stochastic action samples and a trade-off between open-loop consistency and closed-loop reactivity, with existing guidance and test-time search methods being costly or ineffective for sequential control.
Approach
Self-Guidance uses the score predicted from the past observation as the negative term in classifier-free-style guidance, requiring no auxiliary model; Adaptive Chunking keeps executing the queued action chunk and replans only when the cosine similarity between the queue's first action and the new prediction falls below a threshold.
Key results
- Outperforms Vanilla Diffusion Policy by 23.25% and BID by 12.27% on average across six simulated tasks, without extra computational cost (BID reported as 16x FLOPs and 2x latency).
- On Push-T under stochastic perturbation P=1, the method reaches 0.718 vs 0.322 for Vanilla DP and 0.454 for BID.
- In real-world stochastic pick-and-place over 20 trials, the method reaches 70% success, higher than Vanilla DP, EMA and BID; the method runs at about 29 Hz vs about 16 Hz for BID.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Self-Guidance and Adaptive Chunking | Diffusion Policy · Push-T, P=0 | target area coverage: 0.496 ± 0.080 -> 0.819 ± 0.049 | Helped |
| Add Self-Guidance and Adaptive Chunking | Diffusion Policy · Push-T, P=1 | target area coverage: 0.322 ± 0.043 -> 0.718 ± 0.026 | Helped |
| Add Self-Guidance and Adaptive Chunking | Diffusion Policy · Push-T, P=2 | target area coverage: 0.231 ± 0.032 -> 0.413 ± 0.028 | Helped |
| Add Self-Guidance and Adaptive Chunking | Diffusion Policy · Push-T, P=3 | target area coverage: 0.204 ± 0.023 -> 0.261 ± 0.017 | Helped |
| Add Adaptive Chunking (ablation) | Diffusion Policy · Six simulation benchmarks | success rate | Helped |
| Add Self-Guidance (ablation) | Diffusion Policy · Six simulation benchmarks | success rate | Helped |
| Combine Self-Guidance and Adaptive Chunking (ablation) | Diffusion Policy · Six simulation benchmarks | success rate | Helped |
| Use cosine similarity instead of L1 distance in Adaptive Chunking (ablation) | Diffusion Policy · Push-T | target area coverage: 0.621 ± 0.012 -> 0.72 ± 0.037 | Helped |
| Use cosine similarity instead of L2 distance in Adaptive Chunking (ablation) | Diffusion Policy · Push-T | target area coverage: 0.639 ± 0.03 -> 0.72 ± 0.037 | Helped |
| Replace past-state negative guidance with a noised observation (ablation) | Self-Guidance · Simulation tasks | performance | Hurt |
| Use Time-Step Guidance instead of Self-Guidance (ablation) | Diffusion Policy · Simulation tasks | performance | No clear effect |
| Use EMA smoothing with decay rate λ=0.5 | Diffusion Policy · Push-T, P=0 | target area coverage: 0.496 ± 0.080 -> 0.456 ± 0.089 | Hurt |
Surprising findings
- Naive CFG with an unconditional branch performs very poorly on sequential control (near 0.07 success on Push-T in an early table), while using the past state as negative guidance helps.
- Injecting noise into the current observation as the negative term (Noised Observation) performs worse than vanilla.
Limitations
- Adaptive chunking's per-step similarity evaluation makes its computational cost comparable to closed-loop control.
- Gains depend on tuned guidance weight w and similarity threshold tau that vary by task.
- Evaluated mainly on small-scale simulation and a single real-world pick-and-place task.
Experimental setup
Datasets: PushT Robomimic (Square, Lift, Can, Transport) Kitchen LIBERO-Spatial real-world SO-100 pick-and-place demonstrations (300 episodes)
Benchmarks: Push-T Robomimic Square Robomimic Lift Robomimic Can Robomimic Transport Kitchen LIBERO-Spatial real-world SO-100 pick-and-place
Metrics: success rate target area coverage (Push-T)
Baselines: Vanilla Diffusion Policy EMA temporal ensembling BID AutoGuidance CFG Time-Step Guidance Noised Observation guidance
Builds on: Diffusion Policy classifier-free guidance AutoGuidance bidirectional decoding ACT temporal ensembling
Closest papers by meaning
- Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.91)
- One-Step Diffusion Policy: Fast Visuomotor Policies via Diffusion Distillation (ICML 2025 · similarity 0.91)
- Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control (NeurIPS 2026 · similarity 0.90)
- Diffusion Model-Augmented Behavioral Cloning (ICML 2024 · similarity 0.89)
- SkiP: When to Skip and When to Refine for Efficient Robot Manipulation (NeurIPS 2026 · similarity 0.89)
- Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.