AllenAI Releases ToyBlocks Simulation Framework for Robot Learning and Evaluation

• AllenAI (Ai2) released ToyBlocks, an open simulation framework for robot arm and tabletop block manipulation, as the first instance of the FoundationWorlds framework for controlled robot learning.
• Initial data release includes 85,336 long-horizon demonstrations totaling approximately 1,600 hours of robot execution data and 144 million frames.
• Zero-shot sim-to-real transfer experiments on a Franka FR3 arm showed ManiFlow achieving 81.1% real-world success (80.0% in simulation) and a hierarchical Molmo 2 + ManiFlow system achieving 50.0% real-world success (55.5% in simulation).
• Frontier models (Claude Opus 5.5, Claude Fable 5.1) solved 30/30 block-stacking tasks and achieved 63% success on search-and-spell tasks, while low-level VLA models (ManiFlow, π0.5, MolmoAct 2) showed poor generalization beyond training ranges.

Technical Components

• Skills generated via Reinforcement Learning and motion planning in Isaac Lab using privileged simulator information to avoid teleoperation.
• Demonstrations generated by a Task-and-Motion Planner (TAMP) scaling to 24x real-time performance on a single GPU.
• Evaluation metrics include success rates, execution speeds, forces exerted, and task progress.

World Model Evaluation

• Wan 2.2 (5 billion parameters) identified as the most accurate video prediction model for block motion and task outcomes.
• Adding 100 model-generated trajectories from Wan 2.2 raised policy success in an unseen region from 19% to 78%.
• Astra and LeWM models tested for agentic simulation and latent prediction, with LeWM showing drift over full trajectories.