ToyBlocks: A controlled world for robot learning and evaluation

ToyBlocks is a new framework designed to help researchers study robotics in a controlled, simulated environment. This framework is particularly useful for understanding how robots learn and why they fail, as many demonstrations of robot learning are limited in scope and don't explain the underlying mechanisms. FoundationWorlds, the concept behind ToyBlocks, provide a setting for studying questions such as how models improve with more training data and how to combine high-level reasoning with precise physical control.

The first FoundationWorld introduced is ToyBlocks, a simulation framework centered around a robot arm manipulating blocks on a tabletop. While the domain is simple, the robot must still understand and manipulate the blocks using various actions like pushing, flipping, and stacking. ToyBlocks automatically generates demonstrations, which are recordings of the robot completing tasks in simulation, without the need for time-consuming teleoperation.

This means researchers can easily generate more demonstrations and evaluate performance on tasks that were not part of the training set.

ToyBlocks is built around four key requirements: realistic physics and appearance, automatic generation of tasks and demonstrations at scale, challenges that combine reasoning with dexterity, and broad coverage of tasks within a domain. By focusing on a narrow scope, ToyBlocks is able to fulfill all these requirements. Within this constrained setting, researchers can investigate how a robot learns to perform any reasonable task while being confident that the simulated arm and blocks behave similarly to their physical counterparts.

ToyBlocks uses reinforcement learning and motion planning to generate skills for short-horizon tasks. These skills can be trained in Isaac Lab, an open-source framework for robot learning in simulation. The framework can also automatically generate demonstration data using a task-and-motion planner that selects and sequences the skills needed to solve a task and plans the robot's movements between them. This process can generate hundreds of plans in parallel on a single GPU, achieving up to 24 times real-time performance.