Pixel open-sources a low-cost mobile manipulator
Jason Liu and Tony Tao released the robot, its data collection rig and its learning stack together, plus a second paper on whole-body manipulation.
Jason Liu (@jasonjzliu) and Tony Tao (@_tonytao_) have released Pixel, a low-cost mobile manipulator that was designed at the same time as the data collection rig and the learning stack that run on it. It is fully open source. There are two papers, one on the robot and system, and one called Pixel-UMI.
The origin story is the part that will sound familiar to anyone who has tried to buy a research platform. Tao says that around December 2025 he wanted to work on mobile manipulation, but the robots available were either too expensive or not performant enough, so they built their own.
What the two papers cover
The first paper covers Pixel itself, described by Liu as a low-cost mobile manipulator co-designed with its UMI system and learning stack. UMI is a handheld data collection setup, a rig a human holds to record demonstrations that a robot policy can then be trained on, so the hardware and the way you gather data for it were shaped around each other rather than bolted together afterwards.
The second paper, Pixel-UMI, is an improved UMI pipeline for learning environment-aware whole-body manipulation. Tao describes that as whole-body manipulation without robot data, meaning the demonstrations come from the handheld rig rather than from operating the robot itself.
Shown versus claimed
What is on the table right now is two papers and an open-source release of the full stack, not a single long autonomy demo that settles how well the policies hold up outside the lab. The authors are pitching the system, hardware plus pipeline plus learning code, as the contribution.
The reaction from the field has been warm. Arthur Allshire (@arthurallshire) called it the most impressive full stack learning project ever and said he was sure many people would be able to build on it.
For anyone who has been priced out of mobile manipulation research, that last bit is the point. The interesting number here is not a benchmark score, it is how many labs can now copy the whole thing.
probably the most impressive full stack learning project ever, and fully open source
I really wanted to work on mobile manipulation, but robots were either too expensive or not performant enough
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.