Humanoid manipulation requires reusable arm skills, yet demonstrating every possible skill combination is impractical. We present Humanoid Semantic Mixture-of-Experts Diffusion Policy (HS-MoDP), which composes separately demonstrated arm skills through independent semantic routing over a shared expert pool. This design preserves arm-specific skill control while sharing action-generation parameters across arms. A semantic scheduler integrates the learned policy with an existing locomotion controller through alternating manipulation and repositioning stages, extending the effective workspace and enabling skill reuse across location changes. To acquire demonstrations of whole-body contact manipulation, we develop G1Exo, an isomorphic exoskeleton that provides direct joint-level control of arm configurations. Experiments on a Unitree G1 show stronger skill composition than the baselines. G1Exo enables more reliable contact teleoperation than end-effector VR-IK, and its demonstrations improve contact-skill learning across four policy architectures. For HS-MoDP, switching from VR-IK to G1Exo demonstrations raises full cabinet workflow success from 35% to 65%. All project hardware designs and source code will be released as open source after the review process.
Figures & tables
Fig. 1 : G1Exo hardware and joint mapping: (a) wearable exoskeleton; (b) corresponding joints on the Unitree G1; (c) exploded view of a modular rotary joint. Labels 1–7 denote shoulder pitch/roll/yaw, elbow, and wrist roll/pitch/yaw, respectively. G denotes gripper control via the handle trigger.
Fig. 2 : Photographs of whole-body contact manipulation teleoperated with G1Exo. The upper row shows the operator wearing G1Exo, and the lower row shows the corresponding G1 robot configurations: (a) hugging a basket against the chest and (b) holding a basket under the arm.
Fig. 3 : Overview of HS-MoDP. A frozen VLM-based semantic scheduler selects base and arm skills for locomotion and manipulation. Arm-specific skill representations condition the MoE diffusion policy, which generates separate action streams for the two arms. Solid and dashed arrows denote action flow and conditioning, respectively.
Fig. 4 : Autonomous HS-MoDP execution: (a) box pushing with toy placement; (b) four left–right combinations; (c) bimanual basket lifting. Panels (a,c) show chronological sequences; (b) shows separate runs.
Task
Success (%)
π0.5
GR00T
MoDE
Ours
Single-arm skill primitives
R: toy
80
65
60
75
R: block
90
100
75
95
L: box
100
100
100
100
L: pumpkin
85
90
80
95
TABLE I : Skill combination. Single-arm skill primitives and dual-arm combinations were evaluated separately.
Fig. 5 : Autonomous HS-MoDP execution: (a) forearm bottle gathering; (b) right-arm–torso basket cradling; (c) cabinet bottle retrieval with concurrent wrist-assisted door closing. Frames in each row proceed chronologically from left to right.
Fig. 6 : Teleoperation on whole-body contact tasks. (a) Success rates. (b) Successful-trial completion times. Small points show participant means; bars and large points show their mean, and error bars indicate the sample standard deviation across participants.
Task
π0.5
GR00T
MoDE
Ours
G
V
G
V
G
V
G
V
Individual skills
Bottle gathering
100
45
100
55
100
50
100
65
Handleless-basket cradle
80
15
70
10
65
10
85
15
L: retrieval
80
85
65
70
80
75
80
75
R: retrieval
75
80
70
75
75
70
80
70
TABLE II : Autonomous whole-body and mobile manipulation success rates (%) by demonstration source.
Fig. 7 : Ablation on box pushing and toy pick-and-place. Box and Toy denote separately evaluated unimanual skill primitives; Combined denotes their concurrent execution.