Humanoid Locomanipulation Is Easier Than You Think

Shivam Sood1,2, Viswesh Nagaswamy Rajesh1, Somnath Sendhil Kumar1, Parv Kapoor1, Sangkyung Kwak1, Srujan Deolasee1, Guillaume Sartoretti2, Jonathan Huang1

1General Robotics 2National University of Singapore

Work done at General Robotics.

One RL policy learns complete humanoid pick-and-place from task rewards, with a single human demonstration used only as a motion prior.

CoRL 2026 Workshop

Try it yourself

The trained policies run live in your browser, in MuJoCo compiled to WebAssembly. Pick an object, drag the robot, the object or the goal, and press Run. While it runs, drag the object or the robot to push them, or move the goal.

Loading…
0.0 s
Drag to move · drag a ring to turn · Drag the object or robot to push · drag empty space to orbit · scroll to zoom

Waiting to load…

Object and goal stay within 3.5 m of the robot's start.

On the real robot

The box-pickup policy transfers zero-shot to the Unitree G1. When the target switches between two boxes, it redirects its grasp and lifts the newly selected box. It keeps its grasp under sideways and downward pulls, and re-grasps and lifts the box after it is pulled from its hands. None of this is scripted: the reward stage is recomputed from the scene during training, so the policy learns to recover after losing contact.

Two different boxes, one policy (real time).
Recovery after a downward pull (real time).
65.3%OmniContact
reference tracking, hours of mocap
94.2%Ours, box-only
trained on boxes only
95.4%Ours + coverage
wider simulation, same reward and clip

Pick-and-place success on 504 shared scenarios with boxes up to 6 kg. Ours: mean of three seeds.

Abstract

Many recent humanoid loco-manipulation works use reinforcement learning (RL) to track motion references, since long-horizon, contact-rich tasks are hard to specify with rewards and hard to explore. Each new object or placement then needs more motion capture, retargeting, or planning. We find that RL can directly learn a complete loco-manipulation task from task rewards, outperform a state-of-the-art approach, and extend to other object types without per-object rewards or tuning. An object-agnostic staged reward built from contacts and geometry specifies pick-and-place, one human–object clip serves only as a motion prior, and randomized simulation supplies the variation. On the same randomized pick-and-place scenarios, our policy trained only on boxes achieves 99.6% success within its training range, compared with 77.7% for a baseline trained on hours of motion capture. The gap widens outside it in object size, shape, mass, and distance (85.7% versus 45.4%). It also completes 38.2% of episodes with unseen object families, and widening the simulated distribution, without new demonstrations or reward changes, raises this to 69.3%. These findings suggest that generalizable humanoid loco-manipulation need not rely on extensive motion libraries, but can emerge from task-driven RL with minimal demonstration guidance.

Method

(a) Learning: randomized simulation feeds a staged task reward (a staircase over approach, grasp, carry, place and release) and one human clip trains a motion-prior discriminator; task and style critics are mixed 0.9 to 0.1 in PPO for one policy that runs without a reference. (b) Training stages: task (boxes), coverage (+ objects), hardware (+ disturbances).

A staged progress reward

The task is split into six stages, approach, grasp, carry, place, release, placed, read from the scene at every step. The reward is a staircase over the stages with a dense ramp inside each one, so dropping the object simply sends the robot back down the staircase. Every stage is defined from contacts and geometry, so the same reward covers boxes, balls, household objects and chairs without grasp points or per-object tuning.

One clip as a motion prior

The only demonstration is one human clip of lifting, carrying and setting down a box, retargeted to the G1. Adversarial discriminators judge how the robot and the object move, never a trajectory to follow, and a separate critic keeps this style signal at 10% of the update: the prior shapes how the robot moves without deciding what it does.

Comparison with OmniContact

OmniContact plans contact-flow references that a tracking policy trained on hours of motion capture executes. We run its released MuJoCo policy on 504 shared scenarios that vary size, mass, shape, approach and goal placement, with boxes up to 6 kg; 194 of them lie outside our box-only training range. An episode succeeds if the object is carried at least 10 cm above the floor for 0.3 s, then released at rest within 0.35 m of the goal, without a fall.

Success on the 504 scenarios, split by the box-only training range, and on OmniContact's Carry Box benchmark (boxes within 3 m). Ours: mean of three seeds. *Zero-shot.
OmniContactBox-only+ Coverage
Success, all 504 (%)65.394.295.4
in-distribution, 310 (%)77.799.699.9
out-of-distribution, 194 (%)45.485.788.1
Lifted (%)92.194.296.2
Failed after lifting (episodes)13504.0
Falls (episodes)6913.72.7
Carry Box, within 3 m (%)82.897.294.9
Other object families (%)–38.2*69.3

The gap opens after the lift

Both systems lift the object in over 92% of episodes, but OmniContact then fails in 135 of them, while our box-only policy completes every episode in which it lifts the object. The difference comes while carrying, setting down and releasing it.

Beyond the training range

Out of distribution, the box-only policy still succeeds in 85.7% of scenarios, against 45.4%. It handles 16 cm cubes, and 55 cm cubes after the coverage stage; OmniContact fails on the smallest and largest cubes and on 6 kg boxes.

(a) Success in-distribution and out-of-distribution; (b, c) success as cube size and mass vary; (d) success on balls, household objects and chairs, zero-shot and after coverage; for OmniContact, ours box-only and ours plus coverage.
(a) Success in and out of the box-only training distribution. (b, c) Success across cube sizes and masses; shaded: box-only training range; hatched masses were never lifted by OmniContact and are excluded from the shared benchmark (eight scenarios per point). (d) Other object families: the box-only policy zero-shot, and after the coverage stage. Ours: mean and range over three seeds.

Ours (box-only) on the left, OmniContact on the right, in the same physics.

OmniContact
Ours, box-only

Episode outcomes on the 504 scenarios; ours: mean of three seeds.

Coverage grows with simulation, not with data

Trained on boxes alone, the policy already completes 38.2% of 274 episodes with balls, household objects and chairs it has never seen. Widening the simulation to include them, with the same reward and no new demonstrations, raises this to 69.3%, and success on the comparison scenarios to 95.4%.

Coverage stage: boxes, balls and a chair at random placements and goals; three independent episodes per round, unedited.

What learning needs

Stage floors onlyNever lifts

Reaching the next stage is rewarded, but nothing indicates how to get there.

No walking guidanceNever delivers

The policy lifts the object but never brings it to the goal.

No motion priorLearns later

The policy still learns the task, about 1,000 PPO iterations later.

Dense progress and walking guidance carry the learning; the one-clip prior mainly speeds it up.

Fraction of training episodes that lift the object and that complete pick-and-place over PPO iterations, for the full method (three seeds), no motion prior, no walking guidance and stage floors only.
Fraction of training episodes that (a) lift the object and (b) complete pick-and-place over PPO iterations. With only stage floors the policy never lifts; without walking guidance it lifts but fails to deliver; without the motion prior it learns about 1,000 iterations later.
Without the motion prior: the task is still solved, with the motion the reward alone produces.