Search papers, labs, and topics across Lattice.
This paper introduces ACE-EGO-0, a unified framework for Vision-Language-Action (VLA) pretraining that effectively combines egocentric human video data with robotic trajectory data. By developing a scalable video-to-action pipeline and a reliability-aware training objective, the framework addresses the challenges posed by divergent action spaces and supervision quality. The results demonstrate that ACE-EGO-0 significantly enhances joint pretraining and fine-tuning performance, achieving state-of-the-art results on benchmark tasks and facilitating real-world manipulation tasks.
Incorporating egocentric human video data into robotic training can lead to substantial performance gains, with ACE-EGO-0 setting new benchmarks in VLA tasks.
Vision-Language-Action (VLA) models benefit from large-scale and diverse embodied data, yet scaling robot trajectory collection is costly and labor-intensive. Recent advances show that large-scale egocentric human videos provide complementary real-world supervision in pretraining. However, joint training on human and robot data remains challenging due to divergences in action spaces, embodiment structures, temporal dynamics, and supervision quality. We introduce ACE-EGO-0, a unified VLA pretraining framework jointly leveraging heterogeneous data sources. To extract large-scale pretraining supervision from egocentric human videos, we build a scalable egocentric video-to-action pipeline that converts raw human videos into robot-format pseudo-action trajectories. To make these labels comparable with robot demonstrations, ACE-EGO-0 uses a unified action representation based on camera-space actions, morphology conditioning, and time-aligned action chunking. To robustly leverage noisy pseudo-action supervision from egocentric human videos, we formulate a reliability-aware training objective with a human auxiliary loss that concentrates supervision on reliable signals. We instantiate ACE-EGO-0 on 4.53K hours of robot and simulation data, together with 1.48K hours of pseudo-action-labeled egocentric human data. Experiments show that incorporating large-scale human supervision under reliability-aware weighting consistently improves both unified joint pretraining and supervised fine-tuning. ACE-EGO-0 achieves state-of-the-art performance on RoboCasa GR1 TableTop and RoboTwin 2.0, while demonstrating strong transfer to real-world bimanual manipulation.
