A controlled study of cross-embodiment VLA transfer, separating strict zero-shot generalization to unseen robot embodiments from pretrain-exposed transfer, and analyzing the effects of state-action representations, embodiment diversity, auxiliary co-training, and target-embodiment exposure.
A stereo-vision VLA that jointly models semantic and geometric cues for fine-grained spatial perception, improving manipulation accuracy and robustness across large camera-pose variations.
A grasping foundation model pretrained on billion-scale synthetic action data, enabling direct sim-to-real transfer, open-vocabulary generalization, and few-shot adaptation to downstream preferences.
* Equal contribution † Corresponding author
Radiohead
A dazzling spectrum of feelings, and perhaps Radiohead’s greatest musical achievement.
Radiohead
Three long-mythologized songs from the OK Computer era.
