AgiBot's New Robot AI Learns Chores Just by Watching Videos

A team from AgiBot Research has created GE-Act 2.0, an AI system that teaches robots how to handle objects by combining two abilities: imagining what will happen next and figuring out what actions produce a desired outcome. Think of it like a chef who watches a cooking video, pictures the dish halfway done, then reverse-engineers the steps. Most past robot AI relied on pre-trained video generators; this one learns from scratch using manipulation footage and trial data.

What makes this exciting is the scaling effect. When the team trained the model on 300 hours of robot interaction video, it succeeded at tasks only 17% of the time. But when they increased the data to 30,000 hours, success jumped to 44% across 100 different tasks, from grasping objects to more complex manipulation skills. Even tasks the robot never trained on improved by nearly 18 percentage points, showing it can transfer skills between different robot bodies and environments. That is a big deal because it suggests robots can learn broadly from cheap video rather than needing custom programming for every single job.

GE-Act 2.0 also follows instructions even when they contradict what the scene suggests. For example, if a robot is told to place a cup on a plate that is already under another object, it will obey the verbal command rather than the obvious physical arrangement. This matters for real-world usefulness, where a robot needs to listen to you, not just pattern-match.

So why should you care? The same techniques that let this robot succeed in labs are what will eventually power robots in warehouses, kitchens, and homes. We are not there yet — 44% success is still far from reliable — but this is one of the clearest signs that robots are learning faster, cheaper, and with less human hand-holding than before.