InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation

Forum topic · 小凯 · 2026-10-04

Summary

InterEvolve is a framework for test-time evolution in humanoid loco-manipulation, enabling a robot controller to solve tasks it was never trained for by repurposing existing skills, learning from its own attempts, and retaining improvements without retraining. The approach rests on the insight that a broad controller already possesses much of the competence a new task requires, which can be unlocked through an expressive yet executable planning-control interface. InterEvolve combines an object-aware forward-backward (FB) behavioral foundation model, where object residuals on a frozen body prior convert new rewards into loco-manipulation behaviors, with reward programs: staged rewards with completion conditions and tunable constants. An LLM agent revises program structure in-context using execution feedback and a library of verified skill programs, while a numerical optimizer tunes constants; candidates are validated in parallel simulation. Experiments show that hand-designed rewards leave most of the FB model's loco-manipulation capability dormant, while InterEvolve's evolved programs unlock it, sometimes via entirely novel strategies. The method handles diverse tasks, complex scenes, and long-horizon compositions in simulation, and evolved skills run autonomously on a physical Unitree G1 using first-person onboard perception. Paper: arXiv 2610.02196, by Zhuo Lin, Sirui Xu, Liuyu Bian, Yu-Xiong Wang, and Liang-Yan Gui.

Paper Overview

Field: Computer Vision / Robotics
Authors: Zhuo Lin, Sirui Xu, Liuyu Bian, Yu-Xiong Wang, Liang-Yan Gui
Date: 2026-10-01
arXiv: 2610.02196

Summary

The paper studies test-time evolution for humanoid loco-manipulation: solving tasks a controller was never trained for by repurposing its existing skills, improving from its own attempts, and retaining what it learns — all without retraining.

Key insight: a broad controller already holds much of the competence a new task needs. This competence becomes accessible through an interface between planning and control that is expressive enough to specify contact-rich, multi-stage interactions, yet executable and measurable enough that execution feedback can guide planning from experience.

InterEvolve realizes this interface with two components:

1. Object-aware forward-backward (FB) behavioral foundation model — object residuals act on a frozen body prior, so that new rewards about the body or an object can be turned into loco-manipulation behavior at test time. 2. Reward programs — tasks are specified as staged rewards with completion conditions and tunable constants. An LLM agent revises program structure in context using execution feedback and a library of verified skill programs, while a numerical optimizer tunes the constants. Each candidate is validated in parallel simulation, so programs evolve new ways to induce, reuse, and combine the controller's existing locomotion competences.

Findings

Hand-designed rewards leave most of the FB model's loco-manipulation capability dormant; InterEvolve's evolved programs unlock it, sometimes through entirely novel strategies.
The method handles diverse tasks, complex scenes, and long-horizon compositions in simulation.
Evolved skills run autonomously on a physical Unitree G1 based on first-person onboard perception.

---

*Auto-collected on 2026-10-04.*

InterEvolve: Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation

Summary

Paper Overview

Summary

Findings

Tags