Safety-Critical Controller Verification via Sim2Real Gap Quantification

Prithvi Akella, Wyatt Ubellacker, Aaron D. Ames

TL;DR

This paper introduces a novel method to bridge the gap between simulated and real-world robotic systems for safety-critical applications. By quantifying the "sim2real gap" – the difference between a model's predictions and actual system behavior – they enable the synthesis and verification of controllers in simulation with a high probability of safe real-world performance, significantly reducing the need for extensive hardware testing.

What Problem Does This Paper Address?

Developing controllers for safety-critical systems, like autonomous robots, often relies on simulations. However, a fundamental challenge is that simulations are imperfect models of reality. This "sim2real gap" means a controller that performs perfectly in simulation might behave unexpectedly or unsafely on the actual hardware. Traditional approaches require extensive, time-consuming, and potentially risky hardware testing and tuning to ensure safety. This paper tackles the problem of how to quantify this sim2real gap probabilistically and leverage that quantification to guarantee safe controller performance in the real world, even with simplified models.

Methods

The core of the paper's contribution lies in its probabilistic approach to quantifying the sim2real gap and using it for verification.

  • Sim2Real Gap Quantification: They define the sim2real gap as a probabilistic bound on the difference between the true system's evolution and the model's predicted evolution. This is framed as an optimization problem. To solve this intractable problem, they employ a scenario-based approach, drawing samples of system states and inputs to estimate a probabilistic bound on the gap. This results in an "uncertain model" that is guaranteed to encompass the true system's behavior with a high probability.
  • Controller Synthesis and Verification: The synthesized "uncertain model" is then used for controller synthesis and verification within simulation. Controllers designed to be safe and performant on this uncertain model are expected to translate to safe and performant controllers on the real hardware. This is achieved by defining safety metrics (e.g., avoiding collisions, reaching a goal) and using probabilistic verification techniques (drawing from prior work on scenario optimization and risk-aware verification) to ensure the controller meets these metrics with high confidence.

Datasets

Not discussed in the paper. The paper focuses on a methodology and demonstrates it on existing hardware platforms.

Key Insights and Ideas

  • Quantifying Uncertainty is Key: Instead of trying to make perfect models, the paper argues for quantifying their imperfections (the sim2real gap) and explicitly accounting for this uncertainty.
  • Probabilistic Guarantees: The method doesn't offer absolute certainty but rather high probabilistic guarantees. This is a pragmatic approach for real-world systems where perfect prediction is impossible.
  • Simulation as a Reliable Tool: By rigorously quantifying the sim2real gap, simulations can become more trustworthy for safety-critical controller development, reducing reliance on expensive and risky hardware testing.
  • Reduced-Order Models are Viable: The approach works even with simplified "reduced-order" models, which are common in robotics, by augmenting them with a quantified uncertainty.

Key Results

The paper successfully demonstrates its pipeline on two hardware platforms: the Georgia Tech Robotarium and a quadruped robot.

  • Probabilistic Sim2Real Gap: They successfully calculated probabilistic sim2real gaps for both systems, showing that their estimated gap bounds were indeed greater than sampled gaps with high probability (e.g., 99.5% for the Robotarium).
  • Satisfactory Hardware Performance: Controllers synthesized and verified using their method exhibited satisfactory performance on the real hardware without any parameter tuning or additional hardware testing beyond what was needed for gap identification.
  • Robustness to Randomization: The controllers performed well across a variety of randomized test scenarios, demonstrating their robustness. For instance, the Robotarium controllers were tested with 40 additional randomized scenarios, all successful, and the quadruped controllers with 10, also all successful.

Limitations and Future Work

  • Distribution Assumption: The method relies on the ability to draw independent samples from an unknown distribution representing system states and inputs. Crafting this distribution is acknowledged as a significant area for future work.
  • Randomized Synthesis Step: The authors suggest adding a randomized synthesis step to their pipeline to automate the generation of safe, effective controllers, further minimizing hardware testing.
  • Generalizability of Maps: While the paper uses projection maps (Mx) and input extension maps (Mu), the abstract nature of these maps might require further investigation for different system-model pairs.

Summary updated

Abstract

The well-known quote from George Box states that: "All models are wrong, but some are useful." To develop more useful models, we quantify the inaccuracy with which a given model represents a system of interest, so that we may leverage this quantity to facilitate controller synthesis and verification. Specifically, we develop a procedure that identifies a sim2real gap that holds with a minimum probability. Augmenting the nominal model with our identified sim2real gap produces an uncertain model which we prove is an accurate representor of system behavior. We leverage this uncertain model to synthesize and verify a controller in simulation using a probabilistic verification approach. This pipeline produces controllers with an arbitrarily high probability of realizing desired safe behavior on system hardware without requiring hardware testing except for those required for sim2real gap identification. We also showcase our procedure working on two hardware platforms - the Robotarium and a quadruped.
BibTeX
@inproceedings{akella2023safety,
  author = {Prithvi Akella and Wyatt Ubellacker and Aaron D. Ames},
  title = {{Safety-Critical Controller Verification via Sim2Real Gap Quantification}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2023},
  doi = {10.1109/icra48891.2023.10161126},
}
Figure