Research map / Reinforcement learning / Imitation learning
AED: Adaptable Error Detection for Few-shot Imitation Policy
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.
Introduces Adaptable Error Detection (AED) for monitoring few-shot imitation policies in novel environments, a cross-domain benchmark of seven manipulation tasks, and PrObe, a pattern-extractor/LSTM detector over policy features that outperforms four baselines on AUROC and AUPRC.
Problem
Detecting behavior errors online in few-shot imitation policies executing in novel environments, where normal states are unseen, errors produce no notable visual change, and full rollout temporal information is unavailable.
Approach
PrObe detects errors in the space of the monitored FSI policy's own feature embeddings rather than from independent visual encoders: a gated pattern extractor with an L1 sparsity loss, an LSTM pattern-flow generator, and a temporal-aware triplet loss compare the pattern flow to task embeddings from demonstrations.
Key results
- PrObe achieves the highest Top-1 counts (15 of 21 for AUROC and 17 of 21 for AUPRC) across three FSI policies and seven tasks, with up to 40% average performance difference over the worst method.
- In an error-correction pilot on Press Button with DCT, only PrObe improves success rate (82.22 ± 10.17%) when paired with a reset correction policy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use rollout augmentation (RA) during AED training (ablation) | LSTMED, DCTED, and PrObe · Pick & Place | AUROC and AUPRC | Helped |
| Use rollout augmentation (RA) during AED training (ablation) | SVDDED · Pick & Place | AUROC and AUPRC | Mixed |
| Use rollout augmentation (RA) during AED training (ablation) | AED methods monitoring SCAN · Pick & Place | AUROC and AUPRC | Helped |
| Add error-triggered correction using PrObe | DCT policy · Press Button | success rate | Helped |
| Use error-triggered correction with other AED methods | DCT policy · Press Button | success rate | Mixed |
| Use error-triggered correction | DCT policy · Press Button | success rate | No clear effect |
Surprising findings
- Rollout augmentation has a minor or negative impact on the one-class SVM-based SVDDED baseline, while helping temporal methods.
Limitations
- Online AED inference and real-world experiments are limited (discussed in appendix).
- Error correction in novel environments is challenging and gains from the correction policy are minor.
Experimental setup
Datasets: AED benchmark (seven FSI tasks)
Benchmarks: AED benchmark
Metrics: AUROC AUPRC success rate
Baselines: SVDDED TaskEmbED LSTMED DCTED
Builds on: DCRL one-class SVM triplet loss SCAN
Closest papers by meaning
- How to Leverage Diverse Demonstrations in Offline Imitation Learning (ICML 2024 · similarity 0.85)
- Modeling Multiple Normal Action Representations for Error Detection in Procedural Tasks (CVPR 2025 · similarity 0.85)
- Videos are Sample-Efficient Supervisions: Behavior Cloning from Videos via Latent Representations (NeurIPS 2025 · similarity 0.84)
- Efficient Imitation under Misspecification (ICLR 2025 · similarity 0.84)
- Beyond-Expert Performance with Limited Demonstrations: Efficient Imitation Learning with Double Exploration (ICLR 2025 · similarity 0.84)
- Error Detection in Egocentric Procedural Task Videos (CVPR 2024 · similarity 0.83)
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.