Researchers have developed a way for a Unitree G1 humanoid to take on new box-handling tasks by changing its task program while keeping its pretrained movement controller fixed. Across eight simulated task families, the system averaged 86.5% success, compared with 34.6% for an AI-written program whose numeric settings were tuned but whose stages and goals were left unchanged.
The challenge is to get more out of movements a robot already knows. A controller trained on human-object interactions may know how to push or lift, yet still fail when a new task requires those motions in a different order, such as tipping a box onto a new face or moving it through an obstacle.
InterEvolve represents a task as a staged reward program. Each stage describes what the controller should optimize and the condition for moving to the next phase, such as approaching an object, making contact, transporting it or letting go. The controller supplies the movement; the program changes which goal the controller pursues at each stage.
After simulated attempts, a large language model proposes edits to the stages and objectives. A numerical optimizer called CMA-ES adjusts their weights and thresholds, while a fixed verifier checks each candidate across different starting conditions. The verifier scores task constraints independently of the reward program, so rewriting a goal cannot also redefine success. Programs that pass are added to a library for later tasks.
The 86.5% result is an equally weighted average across the eight simulated task families, not a physical-robot success rate. Each family used 64 test scenarios, and the paper averages reported metrics over three evaluation runs. Results varied: the evolved program succeeded in 59.4% of feet-only kicking tests and 71.9% of tests pushing a box around an obstacle.
The researchers also ran simulation-evolved programs for two box tasks, pushing and kicking, on a physical Unitree G1. The robot used a torso-mounted RGB-D camera, while perception and control ran on an off-board GPU workstation. The paper reports demonstrations rather than a repeated real-world success rate.
Search remains a simulation job: each round took 16 to 34 minutes on one GPU, so the system cannot revise its program in real time during physical execution. It also cannot call on movements outside the controller’s existing repertoire. The results show a way to reorganize known skills for new objectives; broader hardware testing would be needed to establish how reliably those programs transfer.
Sources
- InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation, arXiv:2610.02196v1, submitted October 1, 2026.
- Representative image, not the Unitree G1 used in the study: Humanoid Robot at International Exhibitions by Om Umekar, licensed under CC BY-SA 4.0.
