Robots need a testbed before they need more data
Ai2's ToyBlocks gives robot learning a controlled world with 85,336 demonstrations. A compact policy beats generative backbones. And a new method infers hidden interiors from collisions.
- ToyBlocks ships with 85,336 long-horizon demonstrations, and its hierarchical Molmo 2 plus ManiFlow system succeeded in 55.5% of simulated episodes and 50.0% of real-world episodes. m:…8330
- In a preprint, La-JEPA achieved state-of-the-art performance on RoboTwin 2.0 and LIBERO-Plus without embodied pretraining, and control-specific architectures were faster and more efficient than generative-inherited backbones. Self-reported m:…9640
- In a preprint, VTAM achieved an 83.4% success rate on real-robot tasks without tactile pretraining, comparable to a rival model that requires 30k hours of pretraining, while running 6 times faster. Self-reported m:…9644
- Adding 100 model-generated trajectories from Wan 2.2 raised a policy's success rate in an unseen region from 19% to 78% in ToyBlocks experiments. m:…8330
A controlled world
Ai2 has released ToyBlocks, an open-source simulation framework that gives robot learning a controlled, reproducible place to fail. The framework pairs a simulated world with automatic task generation and metrics for measuring robot behavior, and it generates demonstrations automatically instead of requiring people to drive the robot by hand. The framework is built around four requirements: realistic physics, automatic task generation, a combination of reasoning and dexterity, and broad task coverage, and it uses Isaac Lab for skill training. m:…8330
The first numbers show why that separation matters. ManiFlow, the low-level execution policy, succeeded in 80.0% of simulated episodes and 81.1% of real-world ones. The full hierarchical system, Molmo 2 reasoning on top of ManiFlow execution, succeeded in only 55.5% of simulated episodes and 50.0% of real-world ones. The gap between the parts and the whole is the story: the pieces work better than the system they form, and a controlled testbed is what makes that gap visible instead of arguable. m:…8330
ToyBlocks also shows what simulated data can buy. Adding 100 model-generated trajectories from Wan 2.2 raised a policy's success rate in an unseen region from 19% to 78%, and the framework generates data at up to 24 times real-time speed. The initial release includes 85,336 long-horizon demonstrations, with a second release planned with larger datasets, more skills, and integrated RL training. m:…8330
Compact policies beat borrowed backbones
A preprint this week argues robot policies should be built for control, not borrowed from generative models. La-JEPA trains a compact policy transformer from scratch on JEPA features and achieves state-of-the-art performance on RoboTwin 2.0 and LIBERO-Plus without embodied pretraining. Its authors report that JEPA features outperform video VAE latents in robotic control, particularly in out-of-distribution scenarios, and that control-specific architectures are faster and more efficient than generative-inherited backbones. The implication is practical: the expensive part of a robot policy may be the wrong architecture, not the missing data. m:…9640
A second preprint makes the same case for touch. VTAM, a visuo-tactile world-action model, reached an 83.4% success rate on real-robot tasks without any tactile pretraining, close to a rival model that reached 84.5% and requires 30k hours of pretraining. It did so while running 6 times faster than that rival and 16 times faster than another, by avoiding full video denoising at inference time. And the model has a sharp negative finding of its own: feeding tactile features straight to the action expert, bypassing the tactile world model, led to failure on every task tested. The world model is doing the work, not the raw sensor feed. m:…9644
Seeing inside, and simulating ahead
A third preprint tackles a different measurement problem: inferring what is inside an object from how its surface deforms when it collides. The method replaces implicit density updates with explicit structural hypothesis-space reasoning, using gradient-free and differentiable Material Point Method simulations, and outperforms density-based baselines in both hidden-structure reconstruction and future state prediction. Differentiable MPM simulation refines structural hypotheses as Lagrangian particle systems. The application is robotic manipulation and nondestructive inspection, where internal geometry matters and cannot be seen directly. m:…9652
The field is converging on a simple order of operations: measure first, in a world you control, and only then scale. ToyBlocks, La-JEPA, VTAM, and Structure from Collision all point the same way: the bottleneck is not the model, it is the measurement. Simulation is not a shortcut past that problem. It is the instrument that finally measures it. m:…8330
What is still unresolved
- Statistical confidence across tasks for the hierarchical ToyBlocks system requires further trials.
- Generalization of action models to new spatial layouts remains far from solved.
- The scope and robustness of Structure from Collision in complex interior topologies is limited, per the paper.
- No code or dataset release was explicitly mentioned for VTAM, La-JEPA, or Structure from Collision in the material reviewed.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.