Organizations: The authors are with the Department of Mechanical Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong. · The authors are with the College of Information Science and Technology, Eastern Institute of Technology, Ningbo, China. · The author is with the Department of Mechanical Engineering, National University of Singapore, Singapore. · The author is with the Academy of Artificial Intelligence and Advanced Technology, Xi’an Jiaotong-Liverpool University, Suzhou, China. · The author is with the Department of Aeronautical and Aviation Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong.
In recent years, humanoid loco-manipulation has progressed to whole-body grasping, lifting, and transport. A basic requirement underlying these tasks is carrying the object to its destination, which depends on the locomotion controller accurately executing velocity commands from the navigation policy. However, carrying an object in one hand introduces a persistent asymmetric upper-body load that can degrade velocity tracking. Persistent tracking biases can accumulate into position and heading deviations, increasing the risk of collisions and navigation failure. To address this problem, we present ReACT, a framework for accurate velocity tracking under asymmetric upper-body loads. Reflection-consistent disturbance modulation uses the estimated wrench to modulate actor features for load compensation while preserving action correspondence under reflected observations. Cross-load tracking consistency training compares loaded and unloaded conditions under the same velocity commands to promote consistent velocity tracking across loading conditions. Simulation experiments on a Unitree G1 humanoid robot show improved velocity tracking and reduced directional bias under varying asymmetric loads. Real-world experiments demonstrate collision-free goal reaching in all 40 tested navigation trials under unilateral hand loads with a fixed mapless navigation policy.
Figures & tables
Fig. 1: Overview of ReACT.
Component
Dim.
Description
Reflection Transformation
Base angular velocity ωtb
3
(ωx,ωy,ωz)
(−ωx,ωy,−ωz)
Projected gravity gtb
3
(gx,gy,gz)
(gx,−gy,gz)
Velocity command ct
3
(vxcmd,vycmd,ωzcmd)
(vxcmd,−vycmd,−ωzcmd)
Joint position qt−q0
29
Joint position relative to the default configuration
Fq(qt−q0)
Joint velocity q˙t
29
Joint velocity
Fq(q˙t)
Previous full-body command aˉt−1
29
Previous command of the 29 joints
Fq(aˉt−1)
TABLE I: Observation and action components under sagittal reflection.
FALCON
ReACT (Ours)
Protocol
Evx
Evy
Eω
Jτ
Evx
Evy
Eω
Jτ
(m/s)
(m/s)
(rad/s)
(m/s)
(m/s)
(rad/s)
E1: straight
0.1033
0.1212
0.5115
0.03744
0.0632±0.0055
0.0824±0.0032
0.0924±0.0181
0.02236±0.00131
E2: turning
0.1086
0.1401
0.5687
0.03913
0.0579±0.0070
0.0932±0.0100
0.1049±0.0237
0.02389±0.00191
E3: Dynamic A
0.1204
0.1289
0.5574
0.03311
0.0567±0.0015
0.0872±0.0072
0.0877±0.0218
0.01957±0.00160
E3: Dynamic B
0.1184
0.1277
0.5580
0.03292
0.0562±0.0014
0.0867±0.0067
0.0872±0.0212
0.01952±0.00154
TABLE II: Velocity tracking under asymmetric hand loads. ReACT: mean ± sample SD over two training seeds. FALCON: one released checkpoint. Lower is better. Best means are bold.
Fig. 2: Signed directional bias versus unilateral payload mass at vxcmd=1.0 m/s, using an illustrative checkpoint of each method. Negative mass denotes left-hand loading, positive denotes right-hand loading.
Fig. 3: Planar trajectories for straight walking and quarter turns at 1.0 m/s. Top: 2 kg; bottom: 4 kg. Dashed lines show commanded paths; solid lines show measured trajectories from three evaluation trials for an illustrative checkpoint of each method.
E1: straight walking
E2: constant turning
E3: dynamic commands
Method
Evx
Evy
Eω
Jτ
Evx
Evy
Eω
Jτ
Evx
Evy
Eω
Jτ
w/o Consistency
0.2371
0.1747
0.2007
0.02042
0.2823
0.1679
0.2622
0.02117
0.2673
0.1087
0.1622
0.02251
±0.0068
±0.0193
±0.0299
±0.00003
±0.0096
±0.0048
±0.0362
±0.00021
±0.0180
±0.0068
±0.0158
±0.00248
w/o Modulation
0.2338
0.1937
0.2523
0.02087
0.3102
0.2054
0.3135
0.02227
0.2562
0.1611
0.2068
0.02146
±0.0495
±0.0040
±0.0136
±0.00132
±0.0716
±0.0119
±0.0053
±0.00062
±0.0699
±0.0284
±0.0349
±0.00163
ReACT-DR
0.3383
0.1897
0.2204
0.02023
0.3892
0.1757
0.2955
0.01973
0.3677
0.1384
0.1916
0.01808
TABLE III: Component ablations and payload-randomization control. Best means are bold.
Fig. 4: Real-world navigation comparison between ReACT and FALCON under unilateral hand loads of 1.5 kg. S and G denote the start and goal, respectively.
Fig. 5: Upper-body load configurations of the Unitree G1 used in real-world experiments: (a) a 1.5 kg payload in the left hand with the right hand unloaded; (b) a 1.5 kg payload in the right hand with the left hand unloaded.
Method
Evx (m/s)
Evy (m/s)
Eω (rad/s)
ReACT
0.098
0.159
0.525
FALCON
0.202
0.135
0.696
TABLE IV: Real-robot velocity tracking MAE during 1 m/s turning with ∣ωzcmd∣≥0.8 rad/s under 1.5 kg hand loads.
Success
ReACT metrics
Config.
ReACT
FALCON
Path (m)
Time (s)
R1
10/10
0/10
4.17 ± 0.23
11.39 ± 1.69
L1
10/10
0/10
4.36 ± 0.53
14.51 ± 4.32
R2
10/10
0/10
4.63 ± 0.43
27.67 ± 3.80
L2
10/10
0/10
2.91 ± 0.30
10.29 ± 2.72
Total
40/40
0/40
–
–
TABLE V: Real-world navigation under 1.5 kg asymmetric hand loads. Path length and motion time are mean ± standard deviation over successful trials.
Humanoid robots deployed in industrial environments are required to perform load-carrying transportation tasks that tightly couple locomotion and manipulation. However, achieving stable and robust locomotion under varying payloads and upper-body motions is challenging due to dynamic coupling and partial observability. This paper presents a load-aware locomotion framework for industrial humanoids based on a decoupled yet coordinated loco-manipulation architecture. Lower-body locomotion is controlled via a reinforcement learning policy producing residual joint actions on kinematically derived nominal configurations. A kinematics-based locomotion reference with a height-conditioned joint-space offset guides learning, while a history-based state estimator infers base linear velocity and height and encodes residual load- and manipulation-induced disturbances in a compact latent representation. The framework is trained entirely in simulation and deployed on a full-size humanoid robot without fine-tuning. Simulation and real-world experiments demonstrate faster training, accurate height tracking, and stable loco-manipulation. Project page: https://lequn-f.github.io/LALO/
Lequn Fu, Yijun Zhong, Xiao Li +4
Huazhong University of Science and Technology, Wuhan 430074, China · Beijing Humanoid Robot Innovation Center, Beijing 100086, China
Humanoid loco-manipulation of large, heavy objects demands forceful interaction across the entire body. However, such payloads shift a humanoid's center of mass and impose sustained loads across the upper body, challenging balance and command tracking. We present HULK, a whole-body control framework for forceful loco-manipulation. Using model predictive control (MPC) to guide reinforcement learning with predictions of the loaded dynamics, we train two teachers: one tracks arm motions under wrist forces, and the other locomotes while holding large objects against the body. A capture-point control barrier function augments the wrist-force teacher during training to improve balance under load. We distill both teachers into a single policy. Evaluation spans simulation and the Unitree G1. In simulation, the teacher with the barrier function achieves the lowest forward and lateral velocity tracking errors at 10 kg per arm among evaluated controllers and reduces aggregate divergent component of motion (DCM) excursion magnitude by 35.7% relative to MPC-guided reinforcement learning alone. Our wrist-force teacher withstands torso push disturbances of up to 130 N.
An Dang, Arturo Flores Alvarez, Yu-Ming Chen +5
Amazon · University of Michigan · University of California, Los Angeles +1
Many everyday tasks require controlling contact force as well as motion. On a humanoid, this force must often be coordinated across both hands and the whole body, and cannot be specified by kinematic trajectories alone. Therefore, we present FLORE, a framework for learning force-reactive humanoid loco-manipulation. We collect robot-free demonstrations using two force-sensing UMI-style handheld grippers for hand motion and contact wrenches, and wearable trackers for body motion. A force-reactive diffusion policy takes vision, proprioception, and measured wrenches as input, then predicts actions and per-hand force commands. A hybrid compliance whole-body controller with force tracking executes these commands while maintaining balance: an integral of the force error measured by real force/torque (F/T) sensors lets each hand track force accurately along the commanded direction and remain compliant along the others. In simulation, the controller tracks per-hand force commands of varied magnitude and direction, with little loss in whole-body motion accuracy. Across three contact-rich tasks on the robot, the force-reactive policy succeeds in 80% of rollouts, compared with 26.7% for a position-only policy trained on the same demonstrations and executed by the same controller. For more details, please visit our website: https://louis-leee.github.io/FLoRe/.
Sihang Li, Xiaofeng Guo, Jinkun Cao +5
Amazon FAR (Frontier AI & Robotics) · New York University · Carnegie Mellon University +1