RoboForge: Physically Optimized Text-guided Whole-Body Locomotion for Humanoids

TL;DR

RoboForge introduces a physics-optimized, retarget-free framework that couples motion generation and control bidirectionally, improving the physical plausibility and stability of text-guided humanoid locomotion.

Contribution

The paper proposes a novel bidirectional coupling of motion generation and control using a physics plausibility optimization module, eliminating the need for retargeting and enhancing physical realism.

Findings

Improved tracking accuracy and success rates on humanoid robots.

Outperforms retargeting baselines in precision and stability.

Provides a practical, deployable framework for text-guided humanoid locomotion.

Abstract

While generative models have become effective at producing human-like motions from text, transferring these motions to humanoid robots for physical execution remains challenging. Existing pipelines are often limited by retargeting, where kinematic quality is undermined by physical infeasibility, contact-transition errors, and the high cost of real-world dynamical data. We present a unified latent-driven framework that bridges natural language and whole-body humanoid locomotion through a retarget-free, physics-optimized pipeline. Rather than treating generation and control as separate stages, our key insight is to couple them bidirectionally under physical constraints.We introduce a Physical Plausibility Optimization (PP-Opt) module as the coupling interface. In the forward direction, PP-Opt refines a teacher-student distillation policy with a plausibility-centric reward to suppress artifacts such as floating, skating, and penetration. In the backward direction, it converts reward-optimized simulation rollouts into high-quality explicit motion data, which is used to fine-tune the motion generator toward a more physically plausible latent distribution. This bidirectional design forms a self-improving cycle: the generator learns a physically grounded latent space, while the controller learns to execute latent-conditioned behaviors with dynamical integrity.Extensive experiments on the Unitree G1 humanoid show that our bidirectional optimization improves tracking accuracy and success rates. Across IsaacLab and MuJoCo, the implicit latent-driven pipeline consistently outperforms conventional explicit retargeting baselines in both precision and stability. By coupling diffusion-based motion generation with physical plausibility optimization, our framework provides a practical path toward deployable text-guided humanoid intelligence.