A G1 humanoid wrote its own code to fetch a Pringles can
An-Chieh Cheng's team says ASENA lets a coding agent write, run and fix the robot's programs while the task is happening.
An-Chieh Cheng posted ASENA, a system in which a coding agent writes, runs and fixes a robot's programs while it is doing the task. In the clip, a Unitree G1 humanoid finds its way to the pantry, picks up a Pringles can and hands it over.
No pre-built map. No training for this task.
The pitch is in one line from the post. "No pre-built map. No training for this task." So the robot is not following a map it was handed in advance, and nobody trained a policy specifically for pantry snack retrieval. The agent figures out the program as it goes.
The self-evolving part
The second claim is the more interesting one. Cheng says what works gets saved as code for the next task, with no retraining needed. That is a different loop from most humanoid work right now, where improvement means collecting more demonstrations and running another training job. Here the artifact that improves is source code, not model weights.
Xiaolong Wang, summing up the idea, called it robots "writing their own skills, and getting better and cheaper at it the more they practice". Cheaper is doing real work in that sentence. If a saved program is reused instead of regenerated, the agent should burn fewer tokens the second time it faces a similar job.
The work is credited as joint, with Cheng listing @Isabella__Liu, @edmund_bu, Johan Bjorck, @yin_hongxu, @zhengyiluo, @jankautz, @DrJimFan, @yukez and @Sifei30488L.
Shown versus claimed
What is shown is one successful run of one errand. Every structural claim around it, no map, no task-specific training, code that carries over to the next task, comes from the team's own post. Nobody outside the group has run ASENA, and the post does not give a success rate, a number of trials, or a list of other tasks the saved code transferred to.
That makes this a promising demo rather than a result. The fetch itself is not the hard part. Whether the saved-code loop actually compounds across many different tasks is, and the thread does not answer it yet.
Robots writing their own skills, and getting better and cheaper at it the more they practice.
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.