Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

KlaskTron: An Open-Source Platform for Physical Adversarial Multi-Agent RL

Aswin Ramachandran, Jona Schulz, Maurus Derungs, Carlo Angelini, Tobias Meier, Raffaello D'Andrea

November 2026

TL;DR An open-source desktop robot (under $2500) that plays the two-player game KLASK with two CoreXY gantries, plus an Isaac Lab digital twin for adversarial multi-agent RL. A self-play PPO policy trained with a learned actuator model transfers zero-shot to the hardware, but wins only 1 of 10 games against a recreational human.

What they do

Physical multi-agent RL has few cheap, reproducible testbeds. KlaskTron is a desktop robot that plays KLASK, a two-player game in which magnetically coupled strikers knock a ball into the opponent’s goal. Two opposing CoreXY gantries under the board drive the strikers, and an overhead camera tracks the game. The authors release the CAD files, bill of materials, an Isaac Lab digital twin and a self-play PPO baseline that transfers zero-shot to the hardware.

How it works

  • Hardware. Direct-drive brushless motors on ODrive S1 controllers track velocity commands with 0.013 m/s RMSE. A mean acceleration of 6.25 m/s² crosses the board in under a second. The robot costs under $2500 and takes about 20 hours to build.
  • State estimation. An extended Kalman filter tracks the ball and strikers from camera images. It raises its process noise near walls and strikers, where collisions happen.
  • Learned actuator model. A 3x64 MLP predicts the next velocity from the last 10 commands and velocities. It is trained on scripted trajectories recorded on the robot and replaces the idealized PD actuator in simulation.
  • Training. PPO uses a curriculum that decays dense shaping rewards toward sparse goal rewards, plus self-play against a pool of past checkpoints. 20M steps take about 12 minutes on an RTX 4090.

Results

  • Policies trained with PD actuators and domain randomization failed on hardware. With the learned actuator model they transferred zero-shot, even with domain randomization off.
  • No hardware-level failures occurred over more than 30 consecutive games. Two students retrained the baseline independently, and both of their policies played at a comparable level.
  • Against a recreational human, the agent won 1 of 10 games. Games averaged 31.3 s with 12.4 contacts (7.4 blocks, 5.0 strikes).

Limitations

The baseline needed heavy reward shaping and hand-engineered features, and the authors offer it as a starting point rather than an algorithmic result. Offensive shots often missed by 0.5-1.5 cm, partly because the magnetic coupling lags about 0.4 cm on direction changes. The experiments leave out the game’s magnetic obstacle pieces, and the human comparison is a single 10-game series.

Figures

From Figure 1: The KlaskTron robot, with one 2D gantry per player under the game board.
From Figure 1: The KlaskTron robot, with one 2D gantry per player under the game board.
From Figure 3: The Isaac Lab simulation. Each agent observes ball and striker positions and velocities and outputs co...
From Figure 3: The Isaac Lab simulation. Each agent observes ball and striker positions and velocities and outputs continuous velocity commands.
From Figure 5: Validation of the learned actuator model. Solid lines are the real hardware response, dashed lines the...
From Figure 5: Validation of the learned actuator model. Solid lines are the real hardware response, dashed lines the simulated response to the same commands, and shaded bands the spread over 8 training seeds. Global validation RMSE is 0.013 m/s.
Abstract

Progress in robot learning, particularly physical multi-agent reinforcement learning (MARL), is currently challenged by a limited availability of accessible, standardized benchmarks. While simulation-based MARL has produced remarkable emergent behaviors, from coordinated team play to complex tool use, translating these advances to physical systems remains difficult. A key barrier is infrastructural: physical platforms are often prohibitively expensive, require specialized facilities, or lack open mechanisms and designs for reproducibility. We introduce KlaskTron, an open-source, low-cost (<\$2500 USD), desktop-scale robotic testbed for adversarial MARL based on the dynamic tabletop game KLASK. The platform features dual CoreXY gantries with high-torque brushless motors, enabling the high-speed, precise manipulation required for competitive play. Crucially, we provide a complete ecosystem: hardware designs (CAD, BOM), a GPU-native digital twin in NVIDIA Isaac Lab for high-fidelity modeling and simulation, and a validated sim-to-real baseline using a learned neural actuator model. We demonstrate that this baseline enables zero-shot policy transfer, with trained agents exhibiting emergent adversarial behaviors. To ensure full reproducibility of our results, all project assets are released publicly: hardware CAD/BOM and assembly documentation at [https://idscethzurich.github.io/klask_hardware], inference software stack at [https://github.com/IDSCETHZurich/klask_software] and the simulation and MARL setup at [https://github.com/IDSCETHZurich/klask_rl].

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.