KlaskTron-IL: A Contact-Rich Adversarial Imitation Learning Benchmark with a Validated Sim-to-Real Bridge

Nicklas Sohm, Aswin Ramachandran, Raffaello D'Andrea

November 2026

TL;DR An imitation-learning dataset and benchmark on a low-cost robot that plays the table game Klask, with 80k timesteps of human play recorded on the robot and up to 5M from a PPO expert. Behavioral cloning stays about 0.015 goals per episode below the expert from 80k to 5M demonstrations, and the strongest BC policy runs zero-shot on the real robot.

What they do

KlaskTron is an open-hardware robot (under $2,500) for the two-player tabletop game Klask, with an Isaac Lab twin that already transfers PPO policies zero-shot. This paper adds an imitation-learning layer on top: a three-tier demonstration dataset and baselines that are meant to fail in interpretable ways under contact-rich, adversarial dynamics.

How it works

Platform (from prior work). Two CoreXY gantries move magnetic pegs under the board at 50 Hz, and an overhead camera with an EKF gives a 12D state (both pegs and the ball). Policies output 2D velocities. A learned actuator network, identified on the robot within ±0.2 m/s, replaces the PD controller in simulation.
Dataset. Tier 1 is 80,591 timesteps (27 min) of human play recorded on the robot. Tier 2 mirrors it across the board’s symmetry axes (322k). Tier 3 is 5M timesteps of PPO self-play from an expert that itself runs on hardware, used at 80k/250k/1M/5M.
Baselines. An MLP BC policy, GAIL, and PPO with random or BC initialization under sparse or dense reward.

Results

BC plateaus: from 80k to 5M steps it stays about 0.015 goal differential below the expert, while held-out R² rises from 0.44 to 0.67. Trained on human data, its pegs drift to the walls.
GAIL’s discriminator ranks peg-ball proximity perfectly (Spearman ρ = -1.0), but its logit spans only 0.00-0.09 and the policy turns passive.
Under sparse reward, BC initialization cuts peg-in-goal terminations from 34% to 13%; under dense reward it has no detectable effect.
On hardware, BC-5M shifts less from its simulated behavior than the PPO expert (normalized W1 0.05 vs 0.15) but engages the ball 31% less often.

Limitations

State input only, with no image dataset, and the game’s magnetic obstacles are left out. Human action labels are finite-differenced positions, so velocity inputs must be zeroed to avoid leakage. Only one BC policy has been run on hardware. GAIL and BC+RL command up to ±1.0 m/s, five times the actuator model’s range, and real peg speed is about half the simulated value.

Figures

From Figure 1: The physical KlaskTron robot. It exposes the same 12D state / 2D velocity interface as the simulator, ... — From Figure 1: The physical KlaskTron robot. It exposes the same 12D state / 2D velocity interface as the simulator, and a learned actuator model lets a single trained policy run unchanged on either.

From Figure 2: Peg occupancy of a BC rollout (left) vs. human demonstrations (right). The BC policy occupies wall reg... — From Figure 2: Peg occupancy of a BC rollout (left) vs. human demonstrations (right). The BC policy occupies wall regions; the demonstrations do not.

From Figure 3: BC data scaling. Performance sits roughly 0.015 goal differential per episode below the expert at all ... — From Figure 3: BC data scaling. Performance sits roughly 0.015 goal differential per episode below the expert at all four scales, with overlapping 95% CIs (n=5 seeds per scale). 60x more demonstration data does not close the gap.

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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong