Learning Getting-Up Policies for Real-World Humanoid Robots
Xialin He, Runpei Dong, Zixuan Chen, Saurabh Gupta
2025-02-18
Summary
This paper talks about SWE-Lancer, a new way to test how good AI language models are at doing real software engineering jobs. It's like giving AI a bunch of coding tasks that real freelance programmers would do and seeing how much money they could earn.
What's the problem?
Current ways of testing AI's coding skills don't really show how well they can handle real-world programming jobs. It's hard to know if AI is actually ready to do the work of human programmers in practical situations.
What's the solution?
The researchers created SWE-Lancer, which uses over 1,400 real coding tasks from Upwork, a freelancing website. These tasks are worth a total of $1 million in real money. They test the AI on both writing code and making decisions about which code to use. They then check the AI's work using thorough tests that experienced programmers have triple-checked.
Why it matters?
This matters because it gives us a more realistic picture of how close AI is to being able to do the job of human programmers. By linking AI performance to real money earned, we can better understand how AI might impact jobs and the economy in the future. It also helps researchers improve AI coding skills in ways that are actually useful in the real world.
Abstract
Automatic fall recovery is a crucial prerequisite before humanoid robots can be reliably deployed. Hand-designing controllers for getting up is difficult because of the varied configurations a humanoid can end up in after a fall and the challenging terrains humanoid robots are expected to operate on. This paper develops a learning framework to produce controllers that enable humanoid robots to get up from varying configurations on varying terrains. Unlike previous successful applications of humanoid locomotion learning, the getting-up task involves complex contact patterns, which necessitates accurately modeling the collision geometry and sparser rewards. We address these challenges through a two-phase approach that follows a curriculum. The first stage focuses on discovering a good getting-up trajectory under minimal constraints on smoothness or speed / torque limits. The second stage then refines the discovered motions into deployable (i.e. smooth and slow) motions that are robust to variations in initial configuration and terrains. We find these innovations enable a real-world G1 humanoid robot to get up from two main situations that we considered: a) lying face up and b) lying face down, both tested on flat, deformable, slippery surfaces and slopes (e.g., sloppy grass and snowfield). To the best of our knowledge, this is the first successful demonstration of learned getting-up policies for human-sized humanoid robots in the real world. Project page: https://humanoid-getup.github.io/