MorphMimic runs Unitree, AgiBot and Booster humanoids on one brain, a merger nobody signed

Nothing big in the last 24 hours. Here's what led.

MorphMimic: one world model and one policy for three different humanoids

Four days after SuperMimic, which inspired it, MorphMimic shares every weight, world model included, across the Unitree G1, AgiBot X2 Ultra and Booster T2, given each robot's actuation, geometry and mass. In simulation it tracks motions about as well as one specialist per robot, and on a real G1 it dances, walks and follows a VR teleoperator, with code released. All three robots must share the same 29-joint layout, and it hasn't been tried on a robot it wasn't trained on (@_bbelousov).

Amazement at the pace: "the pace of research now is wild. supermimic came out 3 days ago. no code release. now there are already derivative works with full code" (@KyleMorgenstein). SuperMimic's author @qiayuanliao: "First follow-up paper of supermimic, in just 2 days!!!" The open question, from @techniahqrobot: "Would the same policy work on another robot that wasn't included in training?"

Heat: low · 56K views, 10 posts · quiet on Reddit

1. Nebius updates its open-source Physical AI Workbench so a coding agent can run the robot pipeline

▶ Watch the video

Built with NVIDIA, it joins data generation, simulation, training and evaluation behind one command-line tool that a coding agent can drive, on Cosmos 3, Isaac Sim, Isaac Lab, GR00T and SONIC. The first workflow from a new Sereact partnership is a sim-to-real pipeline for returns handling at Zalando. RoboForce, Moon Surgical and Bellboy also use it, and customers report up to 40% less engineering time, it says. Apache 2.0, on GitHub.

Few takes yet beyond @lukas_m_ziegler: "The infrastructure layer of physical AI is being built in Europe."

Heat: low · 5K views, 1 post · quiet on Reddit

2. Audi tests Agile Robots' wheeled humanoids at Ingolstadt

The first task is picking parts from bins and putting them in the right order for vehicle assembly, in logistics that handle about 1 million picks a day. Dassault Systèmes and NVIDIA build a digital twin to test the workflows first. It is still a proof of concept (@CyberRobooo).

@CyberRobooo likes the wheels: "On flat factory floors, robots may not need two legs to move between workstations. What matters is whether they can handle varied parts, keep up with production, and do it at a competitive cost."

Heat: low · 8K views, 3 posts · quiet on Reddit

RoboSimStudio: an agent turns one phone video into a physics twin in under an hour

UC Berkeley and NVIDIA researchers built this open toolkit on the Newton physics engine. From a smartphone video of one arm holding a bar while the other wipes it, with no CAD and no calibrated cameras, Claude and GPT-6 Astra each built a simulation, one in 56 minutes. When the real robot moved a support the simulation had fixed, the agent corrected its model. It matters because a phone video is now enough to build a training simulation, and the real robot's failures fix it.

Generative value classifiers: a robot critic that learns from the same actions as the policy

Alexander Swerdlow and Chelsea Finn's group propose a new kind of Q-function that doesn't need the action as an input. Today a robot's critic is trained separately on sparse rewards, "blind to everything the policy learned from the actions in the pretraining data"; theirs is built to learn from those actions the way the policy does (@chelseabfinn). It matters because a critic that pairs with large pretrained policies is what reinforcement learning on real robots needs.