X-Loco: Towards Generalist Humanoid Locomotion Control via Synergetic Policy Distillation

Dewei Wang; Xinmiao Wang; Chenyun Zhang; Jiyuan Shi; Yingnan Zhao; Chenjia Bai; Xuelong Li

cs.ROMar 4, 2026

X-Loco: Towards Generalist Humanoid Locomotion Control via Synergetic Policy Distillation

Authors: Dewei Wang, Xinmiao Wang, Chenyun Zhang, Jiyuan Shi, Yingnan Zhao, Chenjia Bai, Xuelong Li

Organizations: University of Science and Technology of China · Institute of Artificial Intelligence (TeleAI), China Telecom · Harbin Engineering University

Abstract

While recent advances have demonstrated strong performance in individual humanoid skills such as upright locomotion, fall recovery and whole-body coordination, learning a single policy that masters all these skills remains challenging due to the diverse dynamics and conflicting control objectives involved. To address this, we introduce X-Loco, a framework for training a vision-based generalist humanoid locomotion policy. X-Loco trains multiple oracle specialist policies and adopts a synergetic policy distillation with a case-adaptive specialist selection mechanism, which dynamically leverages multiple specialist policies to guide a vision-based student policy. This design enables the student to acquire a broad spectrum of locomotion skills, ranging from fall recovery to terrain traversal and whole-body coordination skills. To the best of our knowledge, X-Loco is the first framework to demonstrate vision-based humanoid locomotion that jointly integrates upright locomotion, whole-body coordination and fall recovery, while operating solely under velocity commands without relying on reference motions. Experimental results show that X-Loco achieves superior performance, demonstrated by tasks such as fall recovery and terrain traversal. Ablation studies further highlight that our framework effectively leverages specialist expertise and enhances learning efficiency.

Figures & tables

Fig. 2: Overview of X-Loco . (a) X-Loco integrates the capabilities of three specialist policies into a vision-based generalist policy via Synergetic Policy Distillation. (b) X-Loco can perform diverse locomotion skills in the real world.

Fig. 3: Terrains used for training and evaluation.

Method	Locomotion							Whole-Body Coordination			Recovery
	Slope		Pit		Stairs			Hanging Bar	Box		Flat
	$R_{\text{succ}}$	$D_{\text{trav}}$	$R_{\text{succ}}$	$D_{\text{trav}}$	$R_{\text{succ}}$	$D_{\text{trav}}$	$\bar{R}_{\text{succ}}$	$R_{\text{succ}}$	$R_{\text{succ}}$	$\bar{R}_{\text{succ}}$	$R_{\text{succ}}$
BeyondMimic [ 28 ]	-	-	-	-	-	-	-	1.000 $\pm$ .000	0.916 $\pm$ .008	0.958 $\pm$ .004	-
MoRE [ 54 ]	0.992 $\pm$ .008	7.863 $\pm$ .018	0.844 $\pm$ .009	6.833 $\pm$ .114	0.926 $\pm$ .009	7.387 $\pm$ .013	0.921 $\pm$ .009	-	-	-	-
PPO [ 43 ]	0.823 $\pm$ .021	6.674 $\pm$ .152	0.793 $\pm$ .018	6.396 $\pm$ .184	0.781 $\pm$ .025	6.565 $\pm$ .141	0.799 $\pm$ .021	-	-	-	-
AHC [ 69 ]	0.968 $\pm$ .005	7.871 $\pm$ .022	0.403 $\pm$ .031	3.179 $\pm$ .254	0.278 $\pm$ .042	2.524 $\pm$ .311	0.550 $\pm$ .026	-	-	-	1.000 $\pm$ .000

TABLE I: Quantitative comparison of X-Loco against baselines and specialist policies. Bold numbers indicates the best performance other than the specialists, and - denotes that the method failed to complete the corresponding task.

Skill	Obstacle Properties	Ranges	Unit
Locomotion	slope incline	$[15,20]$	∘
	pit obstacle height	$[0.30,0.40]$	m
	stair step height	$[0.10,0.15]$	m
Whole-Body Coordination	obstacle vertical clearance	$[0.87,0.95]$	m
Whole-Body Coordination	obstacle height	$[0.50,0.65]$	m
Recovery	flat ground	-	-

TABLE II: Terrain Configurations for Evaluation

Fig. 4: Top : Testing the generalist policy on hybrid, challenging terrains. Bottom : Extensibility of the X-Loco to include vision-guided (a) lateral sidling and (b) kneeling crawling.

Method	Locomotion	Whole-Body Coordination	Recovery	Average
Method	$\bar{R}_{\text{succ}}$	$\bar{R}_{\text{succ}}$	$R_{\text{succ}}$	$\bar{R}_{\text{succ}}$
Ours w/o CASS	0.903	0.446	1.000	0.783
Ours w/o MoE	0.692	0.561	0.996	0.749
MoE-3	0.628	0.709	1.000	0.779
MoE-2 (Ours)	0.939	0.845	1.000	0.928

TABLE III: Quantitative results of ablation analysis on CASS and MoE architectures.

Fig. 6: Left : (a) fall recovery and platform traversal; (b) resilience to external disturbances; (c) a continuous sequence of recovery, stair climbing, and rolling under an overhead bar. Right : (Top) a failure case where the robot trips due to lack of camera randomization. (Bottom, from left to right) original depth in simulation, noisy depth, and processed real-world depth.

Joint Names	$K_{p}$ (N $\cdot$ m/rad)	$K_{d}$ (N $\cdot$ m $\cdot$ s/rad)	Action Scale
Hip Roll, Knee	$99.10$	$6.31$	$0.35$
Hip Pitch, Hip Yaw, Waist Yaw	$40.18$	$2.56$	$0.55$
Ankle (P/R), Waist (R/P)	$28.50$	$1.81$	$0.44$
Shoulder (P/R/Y), Elbow	$14.25$	$0.91$	$0.44$

TABLE V: PD Controller Parameters and Action Scaling

Domain Randomization	Sampling Distribution	Unit
Static friction	$\mu_{\text{static}}\sim\mathcal{U}[0.6,1.0]$	-
Dynamic friction	$\mu_{\text{dynamic}}\sim\mathcal{U}[0.4,0.8]$	-
Torso payload mass	$m_{\text{payload}}\sim\mathcal{U}[-1.0,5.0]$	kg
Initial joint positions	$\theta_{\text{init}}\sim\mathcal{U}[0.8,1.2]\times\theta_{\text{nominal}}$	-
External push interval	$\Delta t_{\text{push}}\sim\mathcal{U}[5.0,8.0]$	s
Push linear velocity	$v_{x},v_{y}\sim\mathcal{U}[-0.5,0.5],v_{z}\sim\mathcal{U}[-0.2,0.2]$	m/s

TABLE VI: Domain Randomization Parameters. (

\mathcal{U}[\cdot]

: uniform distribution)

Hyperparameter	Specialist Training (PPO)	Generalist Distillation
Number of environments	4096	4096
Learning rate	$1.0\times 10^{-3}$ (Adaptive)	$1.0\times 10^{-3}$ (Fixed)
Num. epochs per iteration	5	8
Num. mini-batches	4	12
Steps per training batch	24	12
Discount factor ( $\gamma$ )	0.99	-

TABLE VII: Optimization Hyperparameters for Specialists and Generalist Policies

Camera Randomization	Value / Range	Unit
Gaussian noise	$\sigma_{\text{noise}}=0.02$	m
Gaussian filter kernel	$k\in\{3\times 3\}$	pixel
Gaussian filter sigma	$\sigma_{\text{filter}}=1.0$	-
Camera position	$\Delta x,\Delta y,\Delta z\sim\mathcal{U}[-0.05,0.05]$	m
Camera rotation	$\theta_{\text{pitch}}\sim\mathcal{U}[-10,5]$ , $\theta_{\text{roll}},\theta_{\text{yaw}}\sim\mathcal{U}[-1,1]$	∘
Camera horizontal FOV	$\Delta\text{FOV}_{\text{h}}\sim\mathcal{U}[-10,10]$	∘

TABLE VIII: Depth Augmentation and Camera Randomization Parameters. (

\mathcal{U}[\cdot]

: uniform distribution)

Fig. 7: Illustration of the case-adaptive specialist selection process based on $b_{t}$ and $I_{t}$ .

Reward Terms	Equation	Weight
Body position	$\exp\left(-\left(\frac{1}{\|\mathcal{B}_{\text{target}}\|}\sum_{b\in\mathcal{B}_{\text{target}}}\\|\mathbf{p}_{b}^{\text{des}}-\mathbf{p}_{b}\\|^{2}\right)/0.3^{2}\right)$	1.0
Body orientation	$\exp\left(-\left(\frac{1}{\|\mathcal{B}_{\text{target}}\|}\sum_{b\in\mathcal{B}_{\text{target}}}\\|\log(R_{b}^{\text{des}}R_{b}^{\top})\\|^{2}\right)/0.4^{2}\right)$	1.0
Body linear velocity	$\exp\left(-\left(\frac{1}{\|\mathcal{B}_{\text{target}}\|}\sum_{b\in\mathcal{B}_{\text{target}}}\\|\mathbf{v}_{b}^{\text{des}}-\mathbf{v}_{b}\\|^{2}\right)/1.0^{2}\right)$	1.0
Body angular velocity	$\exp\left(-\left(\frac{1}{\|\mathcal{B}_{\text{target}}\|}\sum_{b\in\mathcal{B}_{\text{target}}}\\|\boldsymbol{\omega}_{b}^{\text{des}}-\boldsymbol{\omega}_{b}\\|^{2}\right)/3.14^{2}\right)$	1.0
Anchor position	$\exp\left(-\\|\mathbf{p}_{\text{anchor}}^{\text{des}}-\mathbf{p}_{\text{anchor}}\\|^{2}/0.3^{2}\right)$	0.5
Anchor orientation	$\exp\left(-\\|\log(R_{\text{anchor}}^{\text{des}}R_{\text{anchor}}^{\top})\\|^{2}/0.4^{2}\right)$	0.5

TABLE IX: Reward Function Definitions for the Whole-Body Coordination Specialist (

\pi_{w}

Reward Terms	Equation	Weight
Track lin. vel.	$\exp\{-\frac{\\|\bm{v}_{\text{lin}}^{\text{cmd}}-\bm{v}_{\text{lin}}\\|^{2}_{2}}{0.5}\}$	$5.0$
Track ang. vel.	$\exp\{-\frac{(\bm{\omega}_{\text{yaw}}^{\text{cmd}}-\bm{\omega}_{\text{yaw}})^{2}}{0.5}\}$	$5.0$
Joint acc.	$\\|\ddot{\theta}\\|_{2}^{2}$	$-5\times 10^{-7}$
Joint vel.	$\\|\dot{\theta}\\|_{2}^{2}$	$-1\times 10^{-3}$
Action rate	$\\|\bm{a}_{t}-\bm{a}_{t-1}\\|_{2}^{2}$	$-0.03$
Action smoothness	$\\|\bm{a}_{t}-2\bm{a}_{t-1}+\bm{a}_{t-2}\\|_{2}^{2}$	$-0.05$

TABLE X: Reward Function Definitions for the Locomotion Specialist (

\pi_{l}

Reward Terms	Equation	Weight
Task Rewards		$w^{\text{task}}=1.0$
Orientation	$f_{\text{tol}}(-\theta^{z}_{\text{base}},[0.99,\infty],1,0.05)$	$1.0$
Head height	$f_{\text{tol}}(h_{\text{head}},[1,\infty],1,0.1)$	$1.0$
Style Rewards		$w^{\text{style}}=1.0$
Hip joint deviation	$\sum_{\text{hips}}\mathbb{I}(\max\|\theta_{i}\|>0.9\lor\min\|\theta_{i}\|<0.8)$	$-10.0$
Knee deviation	$\sum_{\text{knees}}\mathbb{I}(\max\|\theta_{i}\|>2.85\lor\min\|\theta_{i}\|<-0.06)$	$-0.25$

TABLE XI: Reward Function Definitions for the Recovery Specialist (

\pi_{r}

). The

f_{\text{tol}}

adopts a Gaussian-style formulation, as detailed in [ 20 ] .

Humanoid Robot Locomotion

Humanoid robot locomotion studies how bipedal robots walk, run, recover from disturbances, and traverse varied terrain.

See all 145 papers

HuMBLE: Human Motion-Driven Behavior Learning for Embodied Locomotion
Oct 7, 2026Mike Zhang, Dongho Kang, Kevin Bergamin +19Human-to-Robot Motion Imitation Humanoid Robot Locomotion
RAI Institute, USA · Boston Dynamics, USA · Carnegie Mellon University, USA
MetaWorld-X: Hierarchical World Modeling via VLM-Orchestrated Experts for Humanoid Loco-Manipulation
Mar 9, 2026Yutong Shen, Hangxu Liu, Penghui Liu +6Humanoid Loco-Manipulation Humanoid Robot Control
School of Information Science and Technology, Beijing University of Technology, Beijing, China · School of Information Science and Engineering, Fudan University, Shanghai, China · University of Alberta, Canada +1
Learning Multi-Phase Robot Loco-Manipulation via Phase-Conditioned Policy Optimization
Oct 26, 2026Simone Manni, Annika Wollschläger, Dario Bellicoso +4Robotic Loco-Manipulation Long-Horizon Robotic Manipulation
Robotic Systems Lab (RSL), ETH Zurich, Zurich, Switzerland. · Robotics and AI Institute (RAI), Zurich, Switzerland.

Oct 7, 2026J/K move · Enter open · S save

HuMBLE: Human Motion-Driven Behavior Learning for Embodied Locomotion

Mike Zhang, Dongho Kang, Kevin Bergamin, Nicola Burger+18

RAI Institute, USA · Boston Dynamics, USA · Carnegie Mellon University, USA

Human-to-Robot Motion Imitation Humanoid Robot Locomotion RL for Legged Locomotion Robot Imitation Learning

Open paper AskPDF arXiv

Despite recent advances in humanoid locomotion, controllers optimized for command tracking and robustness tend to produce mechanical gaits, whereas controllers tied to human motion data often fail to generalize to commands outside the data distribution. This work introduces a learning framework that balances these competing objectives to synthesize real-time steerable, robust, and biomimetic locomotion policies from human data. Using an in-house curated locomotion dataset covering diverse speeds and directions, we first learn a natural locomotion prior policy through a teacher-student distillation process. Specifically, we train a full-body reference-conditioned policy with Reinforcement Learning (RL), then distill it into a lightweight prior policy conditioned solely on proprioception and a planar torso-velocity steering command. Next, we fine-tune the prior policy with multi-task RL to expand command coverage and robustness beyond the data distribution, pairing a goal-conditioned task that tracks arbitrary commands with a reference-guided task that tracks the human data as an explicit style regularizer. We validate our framework on three humanoid robots: the Boston Dynamics Atlas R1, Atlas D1, and Unitree G1. Experimental results demonstrate robust performance across real-world scenarios, including direct user-controlled locomotion in indoor and outdoor environments, and integration as the locomotion layer within hierarchical control stacks. Benchmarks against Tabula Rasa RL policies trained without human data and ablation studies confirm that our framework yields a lightweight, deployable policy that reconstructs coordinated whole-body behavior from a steering command, retaining the human gait characteristics while remaining robust and fully steerable.

Abstract

Humanoid Robot Locomotion

Explore similar work

HuMBLE: Human Motion-Driven Behavior Learning for Embodied Locomotion

MetaWorld-X: Hierarchical World Modeling via VLM-Orchestrated Experts for Humanoid Loco-Manipulation

Learning Multi-Phase Robot Loco-Manipulation via Phase-Conditioned Policy Optimization