ModAR is a World Action Model (WAM) that predicts multiple future modalities one at a time—point tracks, DINO features, then depth—before generating actions. We conduct a systematic study training multimodal WAMs from scratch and find that ModAR outperforms existing WAM formulations across data scales. Our 30.1M scratch-trained model even outperforms a 6B video-model-initialized model finetuned on the same data!

3PoinTr learns manipulation from unconstrained human videos: videos where the human demonstrator can act freely rather than mimicking target robot kinematics. 3PoinTr first predicts dense 3D point tracks, and then conditions a closed-loop multitask policy on these tracks.

RoPotter learns to create bowls on a pottery wheel by using structural priors to inform a continuously deforming mesh representation of the clay.

Long-horizon trajectory optimization can be made tractable by decomposing the problem into contact-mode-specific subtasks. AVO bridges these subtasks with a learned value-function ensemble that guides each subtask toward states that are advantageous for future subtasks, and simultaneously reduces online computation.

MapExRL uses global indoor map predictions and a budget-aware RL policy to exploit structure in indoor environments, outperforming greedy algorithms and other prior methods on long-horizon indoor exploration.

SKOOTR is a brittlestar-inspired tripedal robot that explores the advantages of radially symmetric form-factors. SKOOTR is highly stable and maneuverable, and enables several novel forms of locomotion.