Reproduce “Scaling Rough Terrain Locomotion with Automatic Curriculum Reinforcement Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ziming Li, Chenhao Li, Marco Hutter
LP-ACRL samples tasks by signed episodic-reward changes, avoiding a prescribed difficulty order but using a specified, discretized task space. Simulation results reach 80% success across 600 tasks by 1,500 iterations; most baselines do not by 3,000. A distilled ANYmal D reaches up to 3.0 m/s on flat terrain, 2.5 m/s on challenging terrain, and 3.0 rad/s across terrains; maxima are not simultaneous.
Key insights
- An agent's learning progress serves as an effective and scalable signal for automatically guiding curriculum generation in complex, multi-dimensional task spaces.
- Adaptive task sampling based on learning progress allows for efficient exploration and mastery of diverse behaviors, progressively moving from easier to more challenging tasks while revisiting others to consolidate skills.
- Discretizing all task parameters into a unified discrete task space provides a principled way to apply learning progress-based curriculum management to previously unstructured and large-scale locomotion challenges.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.