Research map / Reinforcement learning / Imitation learning
Latent Policy Barrier: Learning Robust Visuomotor Policies by Staying In-Distribution
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 Latent Policy Barrier (LPB), which trains a diffusion policy on limited expert data plus a latent dynamics model trained on policy-checkpoint rollouts, and steers actions at inference to keep latent states near expert embeddings.
Problem
Behavior-cloned visuomotor policies suffer covariate shift, where small deviations compound into failure; existing fixes (DAgger-style human corrections, synthetic augmentation) are labor-intensive, task-specific, or degrade imitation quality.
Approach
Treat the latent embeddings of expert demonstrations as an implicit barrier (inspired by control barrier functions). A dynamics model trained on expert data and automatically collected checkpoint rollouts predicts future latents, and classifier-style gradient guidance on the diffusion denoising process pulls actions toward expert-like latent states at inference time, with no extra human input.
Key results
- With 20% expert demos, LPB reaches 0.65 on Square, 0.85 on Transport, 0.39 on Tool-Hang, 0.65 on Push-T and 0.75 on LIBERO-10, matching or exceeding Filtered BC and CCIL on every task.
- On real-world belt assembly, LPB raises the base BC success rate from 0.55 to 0.75 over 40 rollouts.
- Under action-noise perturbations on Transport, LPB degrades least among compared methods as perturbation probability rises from 0 to 0.4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add latent-space policy steering to the diffusion policy | Diffusion Policy · Transport | success rate: 0.79 ± 0.034 -> 0.85 ± 0.009 | Helped |
| Add latent-space policy steering to the diffusion policy | Diffusion Policy · Tool-Hang | success rate: 0.29 ± 0.009 -> 0.39 ± 0.009 | Helped |
| Add latent-space policy steering to the diffusion policy | Diffusion Policy · Push-T | success rate: 0.59 ± 0.008 -> 0.65 ± 0.012 | Helped |
| Add latent-space policy steering to the diffusion policy | Diffusion Policy · Libero-10 | success rate: 0.71 ± 0.006 -> 0.75 ± 0.038 | Helped |
| Add latent-space policy steering to the diffusion policy | Diffusion Policy · Square | success rate: 0.65 ± 0.009 -> 0.65 ± 0.019 | No clear effect |
| Use policy rollout data rather than noisy demonstrations to train the dynamics model (ablation) | LPB · Tool-Hang | success rate: 0.30 ± 0.031 -> 0.39 ± 0.009 | Helped |
| Use policy rollout data rather than epsilon-greedy exploration data to train the dynamics model (ablation) | LPB · Transport | success rate: 0.71 ± 0.024 -> 0.85 ± 0.009 | Helped |
| Use classifier guidance instead of MPC action sampling (ablation) | LPB · Tool-Hang | success rate: 0.32 ± 0.028 -> 0.39 ± 0.009 | Helped |
| Use classifier guidance instead of direct gradient descent on actions (ablation) | LPB · Transport | success rate: 0.73 ± 0.025 -> 0.85 ± 0.009 | Helped |
| Use the base policy encoder rather than DINOv2 (ablation) | LPB · Square | success rate: 0.61 ± 0.025 -> 0.65 ± 0.019 | Helped |
| Use the base policy encoder rather than a reconstruction-trained encoder (ablation) | LPB · Transport | success rate: 0.68 ± 0.043 -> 0.85 ± 0.009 | Helped |
| Increase rollout data used to train the dynamics model (ablation) | LPB · Tool-Hang | success rate | Helped |
Surprising findings
- CQL obtains zero success on Square, Transport, Tool-Hang and LIBERO-10 with sparse rewards, while Filtered BC gives only marginal gains on Tool-Hang.
Limitations
- Only corrects short-term, local deviations around expert demonstrations; no long-horizon recovery.
- Dynamics model is trained per task, limiting reuse across domains.
- Latent OOD score requires access to expert training data, so it cannot be computed for pretrained policies without paired data.
Experimental setup
Datasets: Push-T Robomimic (Square, Tool-Hang, Transport) LIBERO-10 UMI cup arrangement demonstrations NIST belt assembly demonstrations
Benchmarks: Push-T Robomimic Square Robomimic Tool-Hang Robomimic Transport LIBERO-10 Real-world cup arrangement Real-world belt assembly
Metrics: success rate
Baselines: Expert BC Mixed BC Filtered BC CQL CCIL LPB-MPC LPB-GD
Builds on: Diffusion Policy DAgger Control Barrier Functions CCIL Conservative Q-Learning
Closest papers by meaning
- FreqPolicy: Frequency Autoregressive Visuomotor Policy with Continuous Tokens (NeurIPS 2025 · similarity 0.90)
- DynaMo: In-Domain Dynamics Pretraining for Visuo-Motor Control (NeurIPS 2024 · similarity 0.89)
- Latent Diffusion Planning for Imitation Learning (ICML 2025 · similarity 0.89)
- Act to See, See to Act: Diffusion-Driven Perception-Action Interplay for Adaptive Policies (NeurIPS 2025 · similarity 0.89)
- Robust Policy Learning via Offline Skill Diffusion (AAAI 2024 · similarity 0.87)
- Implicit Drifting Policy: One-Step Action Generation via Conditional Expert Geometry (NeurIPS 2026 · similarity 0.87)
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.