Deployment-time reliability of learned robot policies

Christopher George R Agia, Jeannette Bohg, Marco Pavone, Stanford UniversityStanford 2026 · DOI

TL;DR

This dissertation tackles the critical challenge of making learned robot policies reliable in real-world, unpredictable environments. It introduces three key contributions: a runtime failure detection system (Sentinel), a data-centric interpretability framework (CUPID) to understand how training data influences policy behavior, and a planning framework (STAP) for coordinating learned skills to achieve long-horizon tasks. These advancements are crucial for enabling trustworthy and scalable robot autonomy.

What Problem Does This Paper Address?

As robots move from controlled lab settings to complex, everyday environments, their learned policies face significant challenges. These include open-ended variability, distribution shifts, and compounding errors that can lead to unreliable behavior. While current research has focused on improving policy capabilities, this work addresses the crucial aspect of deployment-time reliability. This means ensuring robots can operate safely and effectively even when encountering situations not perfectly represented in their training data. Specifically, the paper tackles:

  1. Failure Mitigation: Detecting when a policy is about to fail during operation, before unsafe outcomes occur.
  2. Model Interpretability: Understanding why a policy succeeds or fails, which is essential for debugging and improving it.
  3. Long-Horizon Tasks: Composing individual learned behaviors into complex, multi-step tasks where subtle dependencies can lead to failure.

Methods

The dissertation presents three main methodological contributions:

  1. Sentinel (Runtime Failure Detection): This framework monitors learned robot policies in real-time. It uses a two-pronged approach:
  2. STAC (Statistical Temporal Action Consistency): Detects "erratic" failures by measuring how much a policy's action distributions change over time. This is crucial for policies that might exhibit inconsistent or jittery behavior.
  3. Vision-Language Models (VLMs): Detects "task progression" failures by using VLMs to analyze video of the robot's actions and determine if it's making progress towards the task goal, even if its actions are temporally consistent. This leverages the reasoning capabilities of large language models.
  4. CUPID (Curating Performance-Influencing Demonstrations): This framework uses influence functions from data attribution literature to understand how individual training demonstrations impact a policy's performance. It quantifies the "performance influence" of each demonstration, allowing for principled data curation by either filtering out detrimental demonstrations or selecting high-impact ones. This helps in building datasets that truly improve policy performance.
  5. STAP (Sequencing Task-Agnostic Policies): This framework addresses long-horizon tasks by coordinating learned, task-agnostic skills. It uses Q-functions to estimate the success probability of skill sequences and plans action parameters to maximize this probability, ensuring geometric feasibility. This allows robots to tackle tasks not explicitly seen during skill training.
  6. Text2Motion: A hybrid planner that combines a shooting-based strategy (for efficient planning when feasible plans are easily predicted) with a greedy search strategy (as a fallback for complex scenarios). It leverages LLMs for task planning and STAP for geometric feasibility.

Datasets

The paper utilizes several datasets for training and evaluation:

  • RoboMimic benchmark suite: Used for evaluating CUPID on simulated tasks with varying quality demonstrations.
  • Franka FR3 manipulator: Used for real-world experiments with tasks like "Figure-8," "TuckBox," and "Bookshelf" to evaluate CUPID.
  • TableEnv manipulation task suite: A newly introduced benchmark for evaluating Text2Motion, featuring tasks with varying complexities (long-horizon, lifted goals, partial affordance perception).
  • PushT, Close Box, Cover Object, Push Chair domains: Simulated and real-world environments used to evaluate Sentinel.

Key Insights and Ideas

  • Deployment-time reliability is key: The paper argues that true robot reliability emerges not just from better policies, but from complementary mechanisms operating around the policies during deployment.
  • Divide and conquer for failure detection: Sentinel's approach of splitting failure detection into "erratic" and "task progression" categories allows for specialized, more effective detectors.
  • Data quality vs. performance impact: CUPID demonstrates that simply having "high-quality" demonstrations (by human perception) doesn't guarantee better policy performance; it's the causal influence on performance that matters.
  • Task-agnostic skills enable generalization: STAP's ability to coordinate skills trained independently is crucial for tackling novel, long-horizon tasks.
  • LLMs as symbolic planners: Text2Motion shows that LLMs can be effectively integrated with geometric planners to generate and verify feasible plans from natural language instructions.
  • Geometric feasibility is non-negotiable: For long-horizon tasks with dependencies, simply having a symbolically correct plan is insufficient; geometric feasibility must be explicitly ensured.

Key Results

  • Sentinel: Detected 18% more failures than individual monitoring approaches and enabled earlier intervention. It achieved 95% detection accuracy on real-world tasks.
  • CUPID: Significantly improved policy performance, achieving state-of-the-art results on mixed-quality datasets and using less than 50% of the training data. It also demonstrated robustness to spurious correlations.
  • STAP: Significantly improved long-horizon task success, achieving up to four orders of magnitude faster planning than simulation-based methods.
  • Text2Motion: Outperformed prior language-based planning approaches by 69% in success rate on challenging table-top manipulation tasks, achieving an 82% success rate.

Limitations and Future Work

  • Sentinel: The failure categories might not be exhaustive, and the system doesn't provide formal guarantees on detecting failures. Future work could explore more sophisticated detector combinations and failure prediction.
  • CUPID: The influence estimation can be computationally expensive for very large datasets. Future work could explore more efficient estimation techniques and investigate how data properties influence curation effectiveness. There's also potential to actively inform data collectors based on influential demonstration properties.
  • STAP: Scaling to high-dimensional observations (like images) is a challenge. Future work could explore methods for skill acquisition with high-dimensional inputs, combining learned skills with handcrafted ones, and exploring planning objectives beyond geometric feasibility.
  • Text2Motion: LLM inference can be a bottleneck, and the framework currently operates in closed-world settings. Future work could focus on improving LLM efficiency, extending to open-world settings, and integrating more advanced vision-language models.

Summary updated

BibTeX
@inproceedings{agia2026deployment,
  author = {Christopher George R Agia and Jeannette Bohg and Marco Pavone and Stanford University},
  title = {{Deployment-time reliability of learned robot policies}},
  booktitle = {Stanford},
  year = {2026},
  doi = {10.25740/cp016xz0883},
}
Figure