Research map / Vision-language and multimodal / Robot action and navigation
Rethinking Latent Redundancy in Behavior Cloning: An Information Bottleneck Approach for Robot Manipulation
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.
Adds an information-bottleneck regularizer (mutual information between fused input features and latent, estimated with MINE) to behavior cloning policies to reduce latent redundancy, improving success on CortexBench and LIBERO.
Problem
Behavior cloning policies for robot manipulation may learn latent representations containing redundant input information, and existing BC methods lack a theoretical framework to control this redundancy.
Approach
Extends the information bottleneck principle to behavior cloning by penalizing the mutual information between concatenated multimodal input features and the fused latent representation, estimated with a MINE critic, added to the MSE action loss with a Lagrange multiplier beta; the paper also provides generalization-bound-style theoretical arguments.
Key results
- On CortexBench spatial fusion, ResNet+IB improves average success from 73.40% to 78.11%, and VC-1+IB from 57.04% to 59.28%.
- On LIBERO, BC-VILT+IB improves average success from 48.21% to 53.79%, with a 7.66% gain on LIBERO-Goal; BC-MLP+IB improves average from 16.79% to 25.71%.
- Using a 90M-parameter Diffusion Policy head on LIBERO-Long, DP+IB improves success from 78.0 to 84.0.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IB to ResNet | ResNet behavior cloning · CortexBench DMControl | success rate: 74.93 -> 84.94 | Helped |
| Add IB to visual BC backbones | CortexBench behavior cloning methods · CortexBench | success rate | Helped |
| Add IB to BC-VILT | BC-VILT · LIBERO-Goal | success rate: 76.17 -> 83.83 | Helped |
| Add IB to BC-VILT | BC-VILT · LIBERO-Object | success rate: 43.00 -> 52.00 | Helped |
| Add IB to BC-RNN | BC-RNN · LIBERO-Goal | success rate: 15.17 -> 26.00 | Helped |
| Add IB to BC models | LIBERO behavior cloning methods · LIBERO | success rate | Helped |
| Use temporal Transformer fusion instead of spatial fusion or RNN fusion | LIBERO BC policies · LIBERO | average success rate | Helped |
| Add IB to the real-world VC-1 and CogAct policies | Real-world behavior cloning · Real-world pick and put tasks; language-conditioned multi-task setting | success rate | Helped |
| Add IB to Diffusion Policy | Diffusion Policy · LIBERO-Long | success rate: 78.0 -> 84.0 | Helped |
| Use IB with BC-VILT at beta around 1e-4 (ablation) | BC-VILT · LIBERO | success rate | Helped |
| Increase the number of demonstrations from 10-shot to full-shot (ablation) | BC-VILT · LIBERO-Object | success rate: 56.17% -> 43.00% | Hurt |
| Add IB to BC-VILT | BC-VILT · LIBERO-Goal | mutual information I(X, Z): to one-quarter of its original value | Helped |
Surprising findings
- Temporal fusion methods underperform spatial fusion on simple single-task CortexBench settings, attributed to slower loss reduction.
- Gains on LIBERO-Long are marginal for lightweight BC-VILT, which the authors attribute to limited baseline capacity.
Limitations
- Not evaluated on large VLA architectures and alternative policy heads beyond MLP (except a DP experiment on LIBERO-Long).
- Robustness to domain shifts (environment and task configuration changes) is insufficiently studied.
- Gains are smaller on TriFinger and LIBERO-Spatial, and on long-horizon LIBERO-Long with lightweight backbones.
Experimental setup
Datasets: CortexBench LIBERO real-world UR5 demonstrations
Benchmarks: CortexBench LIBERO real-world UR5 tabletop tasks
Metrics: success rate mutual information I(X,Z)
Baselines: ResNet ViT R3M Voltron VC-1 MPI BC-MLP BC-RNN BC-Transformer BC-VILT CogAct Diffusion Policy
Builds on: information bottleneck MINE behavior cloning VC-1 CogAct
Closest papers by meaning
- Videos are Sample-Efficient Supervisions: Behavior Cloning from Videos via Latent Representations (NeurIPS 2025 · similarity 0.90)
- Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning (NeurIPS 2024 · similarity 0.89)
- Diffusion Model-Augmented Behavioral Cloning (ICML 2024 · similarity 0.88)
- Self-Supervised Keyframe Discovery for Horizon-Invariant Behavior Cloning (NeurIPS 2026 · similarity 0.88)
- How to Leverage Diverse Demonstrations in Offline Imitation Learning (ICML 2024 · similarity 0.88)
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.88)
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.