CMU's Pixel hits 87% task progress with no teleoperation

More detail has landed on Pixel, the low-cost wheeled humanoid from a CMU team that earlier open-sourced its hardware and learning stack. The number now attached to it is 87.0% average task progress across five whole-body tasks, according to Lukas Ziegler's writeup. Tasks include reaching into a fridge and packing a canister into a tight space.

Lukas Ziegler

@lukas_m_ziegler

X

Co-design the whole stack, and the human becomes the simplest way to program the robot.

Lukas Ziegler

@lukas_m_ziegler

X

Each optimizes its own piece, and the mismatches between them quietly cap what the whole system can do.

The headline claim is that some of those tasks were learned from human demonstrations alone, with no robot teleoperation (a human driving the robot directly) at all. Instead a person wears a 3D-printed sensorized rig to collect the demos. The robot's cameras and gripper geometry are matched to that rig, so what the human does is supposed to transfer cleanly to the machine.

Co-design is the actual pitch

Ziegler frames the project as a reaction to how household robots usually get built, with separate teams on hardware, data collection, control and learning, each optimizing its own piece. Pixel co-designs all four together. The robot itself is described as a wheeled humanoid with human-aligned kinematics and compliant grippers.

On top sits a hierarchical policy. A high-level vision-language model (a model that reads images and text) plans the subtasks and can talk to you, while a low-level flow-matching policy turns head and wrist camera views into whole-body actions.

Pixel-UMI does the translation

The companion method, Pixel-UMI, is the part Ziegler calls most interesting. It reconstructs the scene from the demonstration's own depth data, then retargets the person's head and hand motion into collision-aware whole-body motion for the robot. That environment awareness is what makes a human reaching into a fridge turn into a robot arm that does not crash into the fridge door.

Shown versus claimed. The 87.0% figure and the no-teleoperation claim both come from the team's own release as relayed by Ziegler, and no independent run has been reported. Task progress is also a softer measure than a binary success rate, so it is worth seeing what partial credit looks like before treating 87.0% as 87 out of 100 completed tasks. Credited on the work are Tony Tao, @JasonJZLiu, @deepakpathak and the CMU team.