A GPU-Parallel Framework for Heterogeneous Multi-Task Reinforcement Learning
Hebero and DGPO train one policy across 40 robot tasks, reaching 90.1% success with 50 demos.
Researchers introduce Hebero, a GPU-parallel Isaac Lab benchmark for training and evaluating one policy across 40 heterogeneous manipulation tasks. Demonstration-Guided Policy Optimization reuses demonstrations as dense tracking rewards and for asymmetric value learning under sparse rewards. Its IW-ABC variant, using 50 demonstrations per task, reaches 90.1% state-input mean success, 7.8 points above FAMO-ABC, and 93.5% in the visual setting. A single simulation-trained policy also performed four tasks on a physical Piper robot.
- Hebero jointly trains and evaluates one policy on 40 Isaac Lab tasks.
- More parallel replicas per task improve success under a fixed time budget.
- IW-ABC reaches 90.1% state success and 93.5% visual success.
- A sim-trained policy completed four tasks on a physical Piper robot.
From the original article
1 min read at huggingface.co
GPU-parallel simulation provides abundant robot interaction, but existing benchmarks rarely combine this scale with heterogeneous manipulation tasks and standardized multi-task RL evaluation. We introduce Hebero (Heterogeneous Benchmark for Robot Learning), a GPU-parallel Isaac Lab benchmark that enables efficient joint training and evaluation of a single policy across all 40 heterogeneous tasks.
By Rui Zhang, Qiwei Wu, Zhengyu Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.