Research map / Reinforcement learning / Visual and representation RL
Failure Prediction at Runtime for Generative Robot Policies
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 FIPER, a runtime failure predictor for diffusion/flow-matching imitation learning policies that fires only when an RND-based observation-embedding OOD score and an action-chunk entropy score both exceed conformally calibrated thresholds, using only successful rollouts.
Problem
Predicting failures of generative imitation learning robot policies early during runtime, without access to failure data, while distinguishing true failures from benign out-of-distribution situations the policy can handle.
Approach
Combines observation-side novelty (random network distillation on frozen policy observation embeddings) with action-side uncertainty (entropy of sampled action chunks estimated by dimension-wise binning in Cartesian space), aggregates both over sliding windows, calibrates thresholds with conformal prediction on successful rollouts, and flags failure only when both indicators fire.
Key results
- FIPER achieves average TWA 0.65 and balanced accuracy 0.78 with average normalized detection time 0.30, outperforming baselines (best baseline TWA 0.60 for logpZO).
- ACE alone reaches TWA 0.75 on Pretzel versus 0.64 for the best baseline PCA-kmeans, and FIPER reaches accuracy 0.96 on PushChair.
- Conjunction (AND) of observation and action scores yields higher TWA and TNR than OR, with FIPER predicting about 91% of failures.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine observation and action warnings with AND rather than OR (ablation) | FIPER · All environments | TWA: 0.51$^{\pm 0.01}$ -> 0.65$^{\pm 0.01}$ | Helped |
| combine observation and action warnings with AND rather than OR (ablation) | FIPER · All environments | Acc.: 0.53$^{\pm 0.01}$ -> 0.78$^{\pm 0.00}$ | Helped |
| use time-varying thresholds instead of CP band thresholds (ablation) | FIPER with AND · All environments | TWA: 0.62$^{\pm 0.00}$ -> 0.65$^{\pm 0.01}$ | Helped |
| use time-varying thresholds instead of CP constant thresholds (ablation) | FIPER with AND · All environments | TWA: 0.61$^{\pm 0.00}$ -> 0.65$^{\pm 0.01}$ | Helped |
| aggregate uncertainty over a sliding window rather than use only the current timestep (ablation) | FIPER · All environments | TWA | Helped |
| aggregate uncertainty over a sliding window rather than accumulate scores over all past timesteps (ablation) | FIPER · All environments | DT | Helped |
| use ACE rather than STAC for action-based failure prediction | Action-based predictor · Sorting, Stacking, Pretzel | Acc. | Helped |
Surprising findings
- Cumulative uncertainty scores mainly detect failures because failed rollouts are longer, rather than predicting them early.
- PCA-kmeans achieves a large mean score gap between success and failure yet has a TNR of only 0.24, so score separation does not guarantee good prediction performance.
Limitations
- Requires collecting successful calibration rollouts and training a separate RND-OE model.
- Runtime cost may grow with very high-dimensional action spaces such as humanoid robots.
- Evaluated only on single-task vision-based IL policies, not large VLAs.
Experimental setup
Datasets: Sorting demonstrations Stacking (D3IL-style) demonstrations PushT rollouts from Agia et al. PushChair rollouts from Agia et al. Pretzel real-world rollouts
Benchmarks: Sorting Stacking PushT Pretzel PushChair
Metrics: balanced accuracy normalized detection time timestep-wise accuracy (TWA) TPR TNR
Baselines: PCA-kmeans logpZO RND-A STAC FAIL-Detect (logpZO)
Builds on: Random network distillation Diffusion Policy ACT STAC conformal prediction functional band (Diquigiovanni et al.)
Closest papers by meaning
- Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.88)
- SAFE: Multitask Failure Detection for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.86)
- AHA: A Vision-Language-Model for Detecting and Reasoning Over Failures in Robotic Manipulation (ICLR 2025 · similarity 0.86)
- Improving Generative Behavior Cloning via Self-Guidance and Adaptive Chunking (NeurIPS 2025 · similarity 0.86)
- GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- PACE: Phase-Aware Chunk Execution for Robot Policies with Action Chunking (NeurIPS 2026 · similarity 0.85)
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.