An AGIBOT presenter explains how GE-Act 2.0 is pretrained and scaled for robot manipulation, and how it performs across unfamiliar objects and settings.
GE-Act 2.0 learns visual prediction and robot actions together using manipulation examples. Its control-oriented autoencoder is designed to combine motion capture, semantic grounding, and efficient compression.
The visual planner learns likely next frames from raw video, including first-person footage of human hands, while an inverse dynamics model learns from action trajectories without captions. The system keeps failed grasps as training examples and uses a one-pass planner that can be trained end to end.
The planner generates several possible outcomes, and the inverse dynamics model estimates the motion associated with both observed and imagined results. The system selects the future whose inferred motion best matches what happened and optimizes using that option.
Tests across 100 tracking tasks covered basic manipulation with unfamiliar objects in new environments. Increasing training data from 300 to 30,000 hours expanded the skills the model could handle. Large-scale pretraining also supported instructions about object color, shape, and orientation, including adversarial prompts.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Thomas, founder and host at AsiaBits, explains how Chinese phone makers and robotics firms are scaling humanoids, expanding into Europe, and influencing manufacturing plans in Japan.
PRO ROBOTS’ presenter reviews humanoid robots and AI systems showcased at WAIC 2026 in China, including demonstrations from AGIBOT, Fourier Intelligence, No Matrix, and others.
Diana Wolf Torres, robotics newsletter writer at Droids and chief evangelist at Lightwheel, discusses physical AI and robotics at NVIDIA GTC 2026 with host Jonathan Stevens on the Thinking Machine podcast.
