HULK — Learning Whole-Body Forceful Loco-Manipulation for Humanoids

Key Findings

MPC reward guidance lets the whole-body controller push and pull heavy objects, including a 300 kg cart.
Capture-Point Constraints improve robustness to force disturbances and balance recovery, measured by DCM excursion.

300 kg

Forceful cart pulling

The whole-body controller pulls a cart loaded to 300 kg on real hardware.

44.6%

Better velocity performance

Lower tracking error while carrying 10 kg per arm in simulation.

35.7%

Better balance recovery

DCM excursion after one-second torso pushes across the 50–100 N sweep in simulation.

Velocity and recovery metrics compare teachers trained with and without capture-point constraints across five simulation evaluation seeds. The 300 kg cart pull is a separate hardware demonstration. Protocol and full results ↗

MPC Reward Guidance

Two specialists learn with model predictive control (MPC) guidance. Their walking rollouts show planned footsteps and predicted body motion under changing loads.

Grasp teacher walking with an opaque orange fridge in front of the torso, with footstep circles and predicted body-motion ellipsoids — Grasp MPC

Fridge carrying under a 0.5 m/s command as payload varies from 0.5 to 2 kg. Footstep targets and predicted body motion are shown at real-time playback speed.

Force teacher walking with the full footstep-circle horizon and predicted body-motion ellipsoids — Wrist-Force MPC

Walking under a 0.3 m/s command with a 40 N total backward load for five seconds, followed by release. Footstep targets and predicted body motion are shown at real-time playback speed.

Simulation demonstrations with different loads. The displayed 2.1-second preview extends the native 0.7-second MPC horizon.

Matched push-recovery comparison: Ours + CBF-RL on the left and Ours baseline on the right — CBF-RL

Capture-point barrier guidance teaches the force teacher to recover balance during training.

Two specialists. One controller.

A wrist-force teacher and a whole-body-grasp teacher distill their behaviors into one student with a shared command interface.

Three-stage training: wrist-force and whole-body-grasp teachers learn with MPC guidance, then label student rollouts through DAgger. Teacher actors and student use deployment observations; teacher critics use privileged state. — Training pipeline The revised manuscript figure. Select the image to inspect it at full size.

Teacher actors already use deployment observations; privileged state is supplied to their training critics. Simulation tables evaluate wrist-force teachers, while the hardware table evaluates the distilled student.

Command interface

Planar velocity · Turning rate · Base height · Torso orientation · Arm joint targets

Force-teacher training rollouts with applied wrist forces in simulation — Training A · Force teacher

Teacher 1 learns to hold arm targets and follow motion commands under sustained wrist forces.

Grasp-teacher walking rollouts from established grasps in simulation — Training B · Grasp teacher

Teacher 2 learns to walk while carrying objects from established whole-body grasps.

Five-controller simulation comparison of maximum push disturbance recovery — Simulation: Push Disturbance Recovery

A side-by-side simulation comparison of disturbance recovery across FALCON, AMO, Thor, Ours, and Ours + DCM-CBF.

Real World Results

We benchmark against AMO, FALCON, and Thor on four real-world tasks: cart pulling, payload carrying, fridge carrying, and canister carrying.

Hardware highlight

HULK humanoid pulling a cart loaded to 300 kg — 300 kg cart pull

HULK pulls a 300 kg cart in the standalone hardware demonstration.

Cart Pulling

Thor performing cart pulling in the supplementary hardware experiment — Thor

AMO performing cart pulling in the supplementary hardware experiment — AMO

Payload Carrying

Thor performing payload carrying in the supplementary hardware experiment — Thor

AMO performing payload carrying in the supplementary hardware experiment — AMO

Fridge Carrying

Thor performing fridge carrying in the supplementary hardware experiment — Thor

AMO performing fridge carrying in the supplementary hardware experiment — AMO

Canister Carrying

Thor performing canister carrying in the supplementary hardware experiment — Thor

AMO performing canister carrying in the supplementary hardware experiment — AMO

Benchmark Results

Carrying and disturbance rejection

This simulation benchmark measures walking-command tracking under 10 kg per arm and recovery from separate torso pushes, with mean ± standard deviation across five evaluation seeds. Our MPC-RL + DCM-CBF teacher has the lowest forward and lateral errors, smallest DCM excursions and peak tilt, and largest push survived, while AMO has the lowest mean turning error.

Swipe the table sideways to see all results.

Carrying and disturbance rejection. Corresponds to Table I in the paper.
Method	Forward errorm/s ↓	Lateral errorm/s ↓	Turn-rate errorrad/s ↓	Signed DCM marginm ↑	Peak tilt · 50 Ndeg ↓	Maximum pushN ↑
Thor	0.263 ± 0.002	0.265 ± 0.003	0.450 ± 0.006	-0.275 ± 0.016	23.17 ± 0.13	80
AMO	0.276 ± 0.002	0.350 ± 0.004	0.196 ± 0.003	-0.705 ± 0.030	9.94 ± 0.74	50
FALCON	0.367 ± 0.012	0.369 ± 0.013	0.392 ± 0.005	-0.329 ± 0.034	7.91 ± 0.23	80
Ours (MPC-RL)	0.242 ± 0.009	0.257 ± 0.004	0.234 ± 0.008	-0.251 ± 0.011	5.90 ± 0.18	110
Ours + DCM-CBF	0.210 ± 0.007	0.142 ± 0.003	0.197 ± 0.009	-0.162 ± 0.012	5.59 ± 0.24	130

Real-world task performance

This Unitree G1 benchmark reports successful trials at a shared task load and the maximum load completed in all three trials of a separate sweep. Our distilled HULK student leads on cart pulling, fridge carrying, and canister carrying, and ties AMO on payload carrying.

Swipe the table sideways to see all results.

Hardware completion counts at fixed task loads and maximum loads in separate sweeps. Corresponds to Table III in the paper.
Method	Cart push/pullTrials at 150 kg		Payload carryTrials at 11 kg		Fridge carryTrials at 6 kg		Canister carryTrials at 5 kg
Method	Completed ↑	Max load · kg ↑	Completed ↑	Max load · kg ↑	Completed ↑	Max load · kg ↑	Completed ↑	Max load · kg ↑
Thor	2/6	130	2/4	7.25	0/4	2	0/4	0.6
AMO	0/6	75	4/4	12.25	2/4	6	0/4	2.6
FALCON	3/6	200	0/4	3.62	0/4	2.5	0/4	0
HULK	5/6	300	4/4	12.25	4/4	7	4/4	6.1

BibTeX

@misc{hulk2027,
  title  = {HULK: Learning Whole-Body Forceful
            Loco-Manipulation for Humanoids},
  author = {Anonymous Authors},
  year   = {2027},
  note   = {Anonymous manuscript}
}

Draft citation. Publication details and authors will be updated.

Whole-body work, in motion

HULK: Learning Whole-Body Forceful Loco-Manipulation for Humanoids

Interactive Demo

Wrist-Force Perturb

Grasping

Loco-manipulation Demos

Fridge transfer

Canister + door

Key Findings

Forceful cart pulling

Better velocity performance

Better balance recovery

MPC Reward Guidance

Grasp MPC

Wrist-Force MPC

MPC optimization for forces

Joint grasp and ground forces

Locomotion forces

CBF-RL

Two specialists. One controller.

Training A · Force teacher

Training B · Grasp teacher

Simulation: Push Disturbance Recovery

Real World Results

300 kg cart pull

Cart Pulling

HULK (Ours)

FALCON

Thor

AMO

Payload Carrying

HULK (Ours)

FALCON

Thor

AMO

Fridge Carrying

HULK (Ours)

FALCON

Thor

AMO

Canister Carrying

HULK (Ours)

FALCON

Thor

AMO

Benchmark Results

Carrying and disturbance rejection

Real-world task performance

BibTeX