CoRL 2024 Demo Recap | Duy P. Nguyen

Live robot demos rarely go exactly as planned, and our CoRL 2024 demo was no exception.

The Gameplay Filters (Nguyen* et al., 2024) demo started with a setback: after the very first demo attempt on Day 1, the robot broke down. What followed was a full day of on-site debugging, tool borrowing from generous neighbors around the venue, and a lot of collective problem-solving under pressure. It was stressful, exhausting, and oddly bonding, all the things that come with running real robots in the wild.

Day 2 told a different story. With the robot back up and running, we were finally able to show the system as intended. The demo went smoothly, conversations with attendees flowed, and the experience became a reminder of why live demos, despite their risks, are so valuable.

This was my first time fixing a quadruped robot entirely from the inside out, under a serious time crunch. It was a huge confidence boost, and an experience I very much hope not to repeat.

This video is a short recap of that journey: the breakdowns, the fixes, the people who stepped in to help, and the small victory of getting a real system to work in a real conference setting. More than anything, it’s a snapshot of the teamwork and persistence behind the scenes of research demos.

Video: CoRL 2024 Demo Recap

Interestingly, after the failure and repair, the robot’s legs were noticeably degraded, with incorrect encoder readings and dampened actuation performance. As shown in the video below, the manufacturer’s built-in controller could no longer stabilize the robot, causing it to fail on its own.

Video: Manufacturer's built-in controller causes robot from falling by itself

Despite this, our Gameplay Filters solution continued to function, demonstrating strong robustness to real-world degradation and adversarial conditions.

Despite the impressive recent advances in learning-based robot control, ensuring robustness to out-of-distribution conditions remains an open challenge. Safety filters can, in principle, keep arbitrary control policies from incurring catastrophic failures by overriding unsafe actions, but existing solutions for complex (e.g., legged) robot dynamics do not span the full motion envelope and instead rely on local, reduced-order models. These filters tend to overly restrict agility and can still fail when perturbed away from nominal conditions. This paper presents the gameplay filter, a new class of predictive safety filter that continually plays out hypothetical matches between its simulation-trained safety strategy and a virtual adversary co-trained to invoke worst-case events and sim-to-real error, and precludes actions that would cause failures down the line. We demonstrate the scalability and robustness of the approach with a first-of-its-kind full-order safety filter for (36-D) quadrupedal dynamics. Physical experiments on two different quadruped platforms demonstrate the superior zero-shot effectiveness of the gameplay filter under large perturbations such as tugging and unmodeled terrain. Experiment videos and open-source software are available online: https://saferobotics.org/research/gameplay-filter

References