KlaskTron platform enables learning robot control from human play
KlaskTron-IL: A Contact-Rich Adversarial Imitation Learning Benchmark with a Validated Sim-to-Real Bridge
Robotics
Summary
Training robots to act in real-world scenarios with lots of contact, like playing games against each other, is hard. The authors built KlaskTron, a physical robot setup where two robots play against each other, and they created a way to teach robots by watching humans play on it. They also made sure that the skills learned in simulation work right away on the real robot without extra tuning. Their data and tests show why some common learning methods struggle in these kinds of interactions and provide a new benchmark for future work.
What this means in practice
- •For robotics engineers: Train and test robot policies that control physical robots in adversarial tasks without additional real-world tuning.
- •For robotic software developers: Use the dataset and benchmark for developing improved imitation learning algorithms for robots interacting in contact-heavy settings.
Authors
Nicklas Sohm, Aswin Ramachandran, Raffaello D'Andrea
Abstract
KlaskTron is an open-hardware physical platform for two-player adversarial reinforcement learning, with a learned actuator model that supports zero-shot sim-to-real transfer of trained policies. Building on this platform, we contribute KlaskTron-IL: a multi-tier imitation learning dataset (80k human demonstration timesteps collected directly on the physical robot, 320k symmetry-augmented timesteps, and synthetic expert trajectories up to 5M steps generated by the zero-shot-validated PPO policy), and diagnostic baselines that expose three characteristic failure modes of standard IL methods under contact-rich adversarial dynamics: BC's data scaling plateau, GAIL's interpretable but weak-gradient discriminator (Spearman $ρ= -1.0$ peg-ball proximity recovery, but a logit span under 0.1), and an asymmetric BC-prior x reward-density interaction. Because the policy interface used throughout this paper is the same one the actuator model bridges to hardware, a policy trained on this benchmark can run unchanged on the physical robot, as we demonstrate zero-shot for the strongest BC baseline, so long as its commands respect the actuator model's $\pm 0.2$ m/s identification range.