Beyond Binary Success: Sample-Efficient and Statistically Rigorous Robot Policy Comparison

Generalist robot manipulation policies are becoming increasingly capable, but are limited in evaluation to a small number of hardware rollouts. This strong resource constraint in real-world testing necessitates both more informative performance measures and reliable and efficient evaluation procedures to properly assess model capabilities and benchmark progress in the field. This work presents a novel framework for robot policy comparison that is sample-efficient, statistically rigorous, and applicable to a broad set of evaluation metrics used in practice. Based on safe, anytime-valid inference (SAVI), our test procedure is sequential, allowing the evaluator to stop early when sufficient statistical evidence has accumulated to reach a decision at a pre-specified level of confidence. Unlike previous work developed for binary success, our unified approach addresses a wide range of informative metrics: from discrete partial credit task progress to continuous measures of episodic reward or trajectory smoothness, spanning both parametric and nonparametric comparison problems. Through extensive validation on simulated and real-world evaluation data, we demonstrate up to 70% reduction in evaluation burden compared to standard batch methods and up to 50% reduction compared to state-of-the-art sequential procedures designed for binary outcomes, with no loss of statistical rigor. Notably, our empirical results show that competing policies can be separated more quickly when using fine-grained task progress than binary success metrics.

Impact of Different Failures on a Robot's Perceived Reliability

Robots fail, potentially leading to a loss in the robot's perceived reliability (PR), a measure correlated with trustworthiness. In this study we examine how various kinds of failures affect the PR of the robot differently, and how this measure recovers without explicit social repair actions by the robot. In a preregistered and controlled online video study, participants were asked to predict a robot's success in a pick-and-place task. We examined manipulation failures (slips), freezing (lapses), and three types of incorrect picked objects or place goals (mistakes). Participants were shown one of 11 videos -- one of five types of failure, one of five types of failure followed by a successful execution in the same video, or a successful execution video. This was followed by two additional successful execution videos. Participants bet money either on the robot or on a coin toss after each video. People's betting patterns along with a qualitative analysis of their survey responses highlight that mistakes are less damaging to PR than slips or lapses, and some mistakes are even perceived as successes. We also see that successes immediately following a failure have the same effect on PR as successes without a preceding failure. Finally, we show that successful executions recover PR after a failure. Our findings highlight which robot failures are in higher need of repair in a human-robot interaction, and how trust could be recovered by robot successes.

A Systematic Study of Data Modalities and Strategies for Co-training Large Behavior Models for Robot Manipulation

Large behavior models have shown strong dexterous manipulation capabilities by extending imitation learning to large-scale training on multi-task robot data, yet their generalization remains limited by the insufficient robot data coverage. To expand this coverage without costly additional data collection, recent work relies on co-training: jointly learning from target robot data and heterogeneous data modalities. However, how different co-training data modalities and strategies affect policy performance remains poorly understood. We present a large-scale empirical study examining five co-training data modalities: standard vision-language data, dense language annotations for robot trajectories, cross-embodiment robot data, human videos, and discrete robot action tokens across single- and multi-phase training strategies. Our study leverages 4,000 hours of robot and human manipulation data and 50M vision-language samples to train vision-language-action policies. We evaluate 89 policies over 58,000 simulation rollouts and 2,835 real-world rollouts. Our results show that co-training with forms of vision-language and cross-embodiment robot data substantially improves generalization to distribution shifts, unseen tasks, and language following, while discrete action token variants yield no significant benefits. Combining effective modalities produces cumulative gains and enables rapid adaptation to unseen long-horizon dexterous tasks via fine-tuning. Training exclusively on robot data degrades the visiolinguistic understanding of the vision-language model backbone, while co-training with effective modalities restores these capabilities. Explicitly conditioning action generation on chain-of-thought traces learned from co-training data does not improve performance in our simulation benchmark. Together, these results provide practical guidance for building scalable generalist robot policies.

CUPID: Curating Data your Robot Loves with Influence Functions

In robot imitation learning, policy performance is tightly coupled with the quality and composition of the demonstration data. Yet, developing a precise understanding of how individual demonstrations contribute to downstream outcomes - such as closed-loop task success or failure - remains a persistent challenge. We propose CUPID, a robot data curation method based on a novel influence function-theoretic formulation for imitation learning policies. Given a set of evaluation rollouts, CUPID estimates the influence of each training demonstration on the policy's expected return. This enables ranking and selection of demonstrations according to their impact on the policy's closed-loop performance. We use CUPID to curate data by 1) filtering out training demonstrations that harm policy performance and 2) subselecting newly collected trajectories that will most improve the policy. Extensive simulated and hardware experiments show that our approach consistently identifies which data drives test-time performance. For example, training with less than 33% of curated data can yield state-of-the-art diffusion policies on the simulated RoboMimic benchmark, with similar gains observed in hardware. Furthermore, hardware experiments show that our method can identify robust strategies under distribution shift, isolate spurious correlations, and even enhance the post-training of generalist robot policies. Additional materials are made available at: https://cupid-curation.github.io.

SAFE: Multitask Failure Detection for Vision-Language-Action Models

While vision-language-action models (VLAs) have shown promising robotic behaviors across a diverse set of manipulation tasks, they achieve limited success rates when deployed on novel tasks out-of-the-box. To allow these policies to safely interact with their environments, we need a failure detector that gives a timely alert such that the robot can stop, backtrack, or ask for help. However, existing failure detectors are trained and tested only on one or a few specific tasks, while VLAs require the detector to generalize and detect failures also in unseen tasks and novel environments. In this paper, we introduce the multitask failure detection problem and propose SAFE, a failure detector for generalist robot policies such as VLAs. We analyze the VLA feature space and find that VLAs have sufficient high-level knowledge about task success and failure, which is generic across different tasks. Based on this insight, we design SAFE to learn from VLA internal features and predict a single scalar indicating the likelihood of task failure. SAFE is trained on both successful and failed rollouts, and is evaluated on unseen tasks. SAFE is compatible with different policy architectures. We test it on OpenVLA, $\pi_0$, and $\pi_0$-FAST in both simulated and real-world environments extensively. We compare SAFE with diverse baselines and show that SAFE achieves state-of-the-art failure detection performance and the best trade-off between accuracy and detection time using conformal prediction. More qualitative results can be found at https://vla-safe.github.io/.

STITCH-OPE: Trajectory Stitching with Guided Diffusion for Off-Policy Evaluation

Off-policy evaluation (OPE) estimates the performance of a target policy using offline data collected from a behavior policy, and is crucial in domains such as robotics or healthcare where direct interaction with the environment is costly or unsafe. Existing OPE methods are ineffective for high-dimensional, long-horizon problems, due to exponential blow-ups in variance from importance weighting or compounding errors from learned dynamics models. To address these challenges, we propose STITCH-OPE, a model-based generative framework that leverages denoising diffusion for long-horizon OPE in high-dimensional state and action spaces. Starting with a diffusion model pre-trained on the behavior data, STITCH-OPE generates synthetic trajectories from the target policy by guiding the denoising process using the score function of the target policy. STITCH-OPE proposes two technical innovations that make it advantageous for OPE: (1) prevents over-regularization by subtracting the score of the behavior policy during guidance, and (2) generates long-horizon trajectories by stitching partial trajectories together end-to-end. We provide a theoretical guarantee that under mild assumptions, these modifications result in an exponential reduction in variance versus long-horizon trajectory diffusion. Experiments on the D4RL and OpenAI Gym benchmarks show substantial improvement in mean squared error, correlation, and regret metrics compared to state-of-the-art OPE methods.

Is Your Imitation Learning Policy Better than Mine? Policy Comparison with Near-Optimal Stopping

Imitation learning has enabled robots to perform complex, long-horizon tasks in challenging dexterous manipulation settings. As new methods are developed, they must be rigorously evaluated and compared against corresponding baselines through repeated evaluation trials. However, policy comparison is fundamentally constrained by a small feasible sample size (e.g., 10 or 50) due to significant human effort and limited inference throughput of policies. This paper proposes a novel statistical framework for rigorously comparing two policies in the small sample size regime. Prior work in statistical policy comparison relies on batch testing, which requires a fixed, pre-determined number of trials and lacks flexibility in adapting the sample size to the observed evaluation data. Furthermore, extending the test with additional trials risks inducing inadvertent p-hacking, undermining statistical assurances. In contrast, our proposed statistical test is sequential, allowing researchers to decide whether or not to run more trials based on intermediate results. This adaptively tailors the number of trials to the difficulty of the underlying comparison, saving significant time and effort without sacrificing probabilistic correctness. Extensive numerical simulation and real-world robot manipulation experiments show that our test achieves near-optimal stopping, letting researchers stop evaluation and make a decision in a near-minimal number of trials. Specifically, it reduces the number of evaluation trials by up to 40% as compared to state-of-the-art baselines, while preserving the probabilistic correctness and statistical power of the comparison. Moreover, our method is strongest in the most challenging comparison instances (requiring the most evaluation trials); in a multi-task comparison scenario, we save the evaluator more than 200 simulation rollouts.

Can We Detect Failures Without Failure Data? Uncertainty-Aware Runtime Failure Detection for Imitation Learning Policies

Recent years have witnessed impressive robotic manipulation systems driven by advances in imitation learning and generative modeling, such as diffusion- and flow-based approaches. As robot policy performance increases, so does the complexity and time horizon of achievable tasks, inducing unexpected and diverse failure modes that are difficult to predict a priori. To enable trustworthy policy deployment in safety-critical human environments, reliable runtime failure detection becomes important during policy inference. However, most existing failure detection approaches rely on prior knowledge of failure modes and require failure data during training, which imposes a significant challenge in practicality and scalability. In response to these limitations, we present FAIL-Detect, a modular two-stage approach for failure detection in imitation learning-based robotic manipulation. To accurately identify failures from successful training data alone, we frame the problem as sequential out-of-distribution (OOD) detection. We first distill policy inputs and outputs into scalar signals that correlate with policy failures and capture epistemic uncertainty. FAIL-Detect then employs conformal prediction (CP) as a versatile framework for uncertainty quantification with statistical guarantees. Empirically, we thoroughly investigate both learned and post-hoc scalar signal candidates on diverse robotic manipulation tasks. Our experiments show learned signals to be mostly consistently effective, particularly when using our novel flow-based density estimator. Furthermore, our method detects failures more accurately and faster than state-of-the-art (SOTA) failure detection baselines. These results highlight the potential of FAIL-Detect to enhance the safety and reliability of imitation learning-based robotic systems as they progress toward real-world deployment.

How Generalizable Is My Behavior Cloning Policy? A Statistical Approach to Trustworthy Performance Evaluation

With the rise of stochastic generative models in robot policy learning, end-to-end visuomotor policies are increasingly successful at solving complex tasks by learning from human demonstrations. Nevertheless, since real-world evaluation costs afford users only a small number of policy rollouts, it remains a challenge to accurately gauge the performance of such policies. This is exacerbated by distribution shifts causing unpredictable changes in performance during deployment. To rigorously evaluate behavior cloning policies, we present a framework that provides a tight lower-bound on robot performance in an arbitrary environment, using a minimal number of experimental policy rollouts. Notably, by applying the standard stochastic ordering to robot performance distributions, we provide a worst-case bound on the entire distribution of performance (via bounds on the cumulative distribution function) for a given task. We build upon established statistical results to ensure that the bounds hold with a user-specified confidence level and tightness, and are constructed from as few policy rollouts as possible. In experiments we evaluate policies for visuomotor manipulation in both simulation and hardware. Specifically, we (i) empirically validate the guarantees of the bounds in simulated manipulation settings, (ii) find the degree to which a learned policy deployed on hardware generalizes to new real-world environments, and (iii) rigorously compare two policies tested in out-of-distribution settings. Our experimental data, code, and implementation of confidence bounds are open-source.

Residual Q-Learning: Offline and Online Policy Customization without Value

Imitation Learning (IL) is a widely used framework for learning imitative behavior from demonstrations. It is especially appealing for solving complex real-world tasks where handcrafting reward function is difficult, or when the goal is to mimic human expert behavior. However, the learned imitative policy can only follow the behavior in the demonstration. When applying the imitative policy, we may need to customize the policy behavior to meet different requirements coming from diverse downstream tasks. Meanwhile, we still want the customized policy to maintain its imitative nature. To this end, we formulate a new problem setting called policy customization. It defines the learning task as training a policy that inherits the characteristics of the prior policy while satisfying some additional requirements imposed by a target downstream task. We propose a novel and principled approach to interpret and determine the trade-off between the two task objectives. Specifically, we formulate the customization problem as a Markov Decision Process (MDP) with a reward function that combines 1) the inherent reward of the demonstration; and 2) the add-on reward specified by the downstream task. We propose a novel framework, Residual Q-learning, which can solve the formulated MDP by leveraging the prior policy without knowing the inherent reward or value function of the prior policy. We derive a family of residual Q-learning algorithms that can realize offline and online policy customization, and show that the proposed algorithms can effectively accomplish policy customization tasks in various environments.