Event
EXE: Cross-Embodiment Egocentric Video Pretraining for VLA
1 reports · 1 sources
Qingyuan Cai and other authors propose EXE, a cross-embodiment pretraining framework that learns factorized latent action representations from first-person human manipulation videos, separating shared manipulation priors from embodiment-specific signals. The authors claim that this shared action space enables robot policies to acquire generalizable manipulation semantics from diverse human demonstrations, and that the same pretraining applies to both parallel grippers and dexterous hands. The authors claim that on RoboTwin 2.0, RoboCasa GR1, and real manipulation tasks, EXE achieves consistent improvements over no-pretraining baselines.
Report timeline
- EXE: Egocentric Cross-Embodiment Pretraining with Factorized Latent Actions for VLA Models
EXE is an egocentric cross-embodiment pretraining framework that learns factorized latent action representations from human manipulation videos, separating shared manipulation priors from embodiment-specific signals. The authors state that this shared action space lets robot policies acquire generalizable manipulation semantics from diverse human demonstrations, and that the same pretraining can be used for both parallel grippers and dexterous hands. On RoboTwin 2.0, RoboCasa GR1, and real-world manipulation tasks, EXE achieves consistent improvements over baselines without pretraining.
arXiv cs.RO ·