Reliable and Scalable Robot Policy Evaluation with Imperfect Simulators

Abstract

Rapid progress in robot manipulation driven by imitation learning, foundation models, and large-scale datasets has enabled generalization to a wide-range of tasks and environments. However, rigorous evaluation of these policies remains a challenge. Typically in practice, robot policies are often evaluated on a small number of hardware trials without any statistical assurances. We present SureSim, a framework to augment large-scale simulation with relatively small-scale real-world testing to provide reliable inferences on the real-world performance of a policy. Our key idea is to formalize the problem of combining real and simulation evaluations as a prediction-powered inference problem, in which a small number of paired real and simulation evaluations are used to rectify bias in large-scale simulation. We then leverage non-asymptotic mean estimation algorithms to provide confidence intervals on policy performance. Using physics-based simulation, we evaluate both diffusion policy and multi-task fine-tuned \(\pi_0\) on a joint distribution of objects and initial conditions, and find that our approach saves over \(20\%\) of hardware evaluation effort to achieve similar bounds on policy performance.

Approach: Augmenting Real Tests with Simulation via Prediction Powered Inference

We would like to estimate expected performance of a policy \(\pi\) according to a bounded evaluation metric \(M\) over some distribution of tasks and environments \(\mathcal{D}_{\text{env}}\). While gold-standard, tackling this problem using only real-world evaluations is often expensive and time-consuming. On the other hand, simulation evaluations are relatively easier to scale, but can be biased, leading to inaccurate estimates of real-world performance.

This motivates the central question of our work: Can we make valid inferences on the real-world performance of a policy by combining a small amount of real evaluations with a large number of simulations? To answer this question, we present SureSim: Scalable and Reliable Evaluation with Simulation.

Paired evaluations

We set up a real2sim pipeline to collect a small number of paired evaluations in both real and simulation environments. This paired dataset allows us to quantify the bias in simulation evaluations and rectify predictions from large-scale simulations. Importantly, the rectifier is related to the correlation between the paired real and simulation evaluations. In a sense, it quantifies the real-simulation gap and provides a lower bound on the advantage (e.g., minimizing confidence interval width, saving hardware evaluations) we can expect from scaling simulation. If the rectifier variance is smaller than the variance in real evaluations, we can expect to save on real-world evaluations.

Additional simulation evaluations

Relative to the size of the paired dataset, we collect a large number of additional simulations. By running large-scale simulation tests, we can further improve the accuracy of our performance estimates.

Real2Sim Experiments

Using our framework, we evaluate diffusion policy and a multi-task fine-tuned \(\pi_0\) on a joint distribution of tasks and initial conditions. We choose the pick-and-place family of tasks, parameterized by different object types.

SureSim requires fewer hardware evaluations

Compared to obtaining confidence intervals with real-world only trials, SureSim leverages simulation effectively, requiring fewer hardware evaluations to achieve similar bounds. SureSim saves 20 - 25% hardware evaluations to get similar bounds on policy performance.

SureSim illustrates the advantage of scaling simulation evaluations

Instead of scaling up hardware evaluations, we can scale up simulation evaluations up to the limit of the real-simulation gap.