ToyBlocks: A controlled world for robot learning and evaluation
Reported by Hugging Face and Unite.AI2 sources ↓
Hugging Face introduced ToyBlocks, an open simulation framework for studying robot learning with a robot arm and tabletop blocks. The company says it automatically generates training demonstrations and tasks, and can run up to 24 times real time on a single GPU.
Key points
- ToyBlocks generates robot skills and task demonstrations in simulation without human teleoperation
- Researchers can vary training data and evaluate models on tasks excluded from training
- The framework uses reinforcement learning, motion planning and task-and-motion planning
- Hugging Face says hundreds of plans can run in parallel on one GPU, reaching up to 24 times real time
Sources
Details
ToyBlocks is the first of Hugging Face’s planned FoundationWorlds, controlled simulated environments intended to help researchers study robot learning. It centers on a simulated robot arm and tabletop blocks, with tasks including pushing, flipping, stacking, searching and sorting.
The framework uses reinforcement learning and motion planning to generate short robot skills. A task-and-motion planner selects and sequences those skills to complete tasks, then runs the plans in simulation to produce demonstrations labeled by task or skill. Hugging Face says this process can generate and execute hundreds of plans in parallel on one GPU, at up to 24 times real time.
Researchers can adjust the generated training experiences, including emphasizing particular skills, and test models on tasks held out from training. ToyBlocks also varies properties such as block size and table height to test generalization, and trains models to map sensor observations to actions.
Written from coverage as of
Images
- Hugging Face
- Hugging Face
- Hugging Face
- Hugging Face