Researchers have reported a humanoid controller designed to tune two responses to physical contact independently: how far the robot yields to a push and how much it turns under a twist. In simulations, higher commanded linear stiffness reduced contact-link displacement while attitude change stayed nearly constant. Higher angular stiffness reduced attitude change while preserving the position response.
The system, called LAC, combines virtual admittance, a model for how motion should respond to an external push or twist, with whole-body inverse kinematics, which coordinates the robot's joint motion. It synthesizes compliant motions, then trains a single policy using teacher-student reinforcement learning. The work was evaluated in simulation and on a real 23-degree-of-freedom Unitree G1.
A training set built around contact
The training material came from retargeted human-object interaction clips in OMOMO and human-human interaction clips in Inter-X. The researchers retained frames where at least one of 11 upper-body links touched an object or partner robot, creating the contact situations used for augmentation.
The accepted set contained 378,051 episodes, each 10 seconds long, totaling about 1,050 hours. It included 278,258 force-event episodes and 99,793 couple-event episodes. Force events reached all 11 upper-body links, while couple events were limited to arm links; the two hands and torso made up 88% of events.
Reinforcement-learning training ran in Isaac Lab on four GPUs, with 8,192 environments per GPU and 32,768 in total. The two training stages ran for 20,000 and 10,000 iterations, respectively.
Separate dials for pushes and twists
The linear test used a 20-newton peak hand force in three simulated cases. As commanded linear stiffness rose from 10 to 500 newtons per metre, displacement fell monotonically. The contact link's attitude change remained nearly constant across the sweep.
Arm tests showed the corresponding angular response across 10 to 100 newton-metres per radian. In those tests, attitude change decreased monotonically while the position response was preserved. In the 4-newton-metre couple test, attitude change fell from 60 degrees at stiffness 10 to 17 degrees at stiffness 100.
In a MuJoCo comparison with baseline controllers, LAC was the only method with a monotonic displacement decrease in all three reported scenarios. Torso displacement fell from 49 centimetres at the low setting to 8 centimetres at the high setting, with the values averaged over three trials.
From simulator to the physical robot
A real lateral-pull replication followed the simulated response in both trend and range. The supplied results do not give a repeated-trial summary or uncertainty estimate for that check.
On the physical G1, force interactions at the hand, elbow, shoulder and torso produced coordinated adjustments in the arms, base, knees and body height.
A twist changed object handling
On the physical robot, left-palm attitude change was 84 degrees at angular stiffness 10 newton-metres per radian and 15 degrees at 100 newton-metres per radian. A 4-kilogram object fell at the low setting but was carried steadily at the high setting.
VR teleoperation paired low stiffness with soft-object transport and high stiffness with force-demanding tasks. The reported demonstrations included transporting a yoga ball, carrying a heavy box and opening a spring-loaded door.
What remains open
The quantitative comparison averaged three trials and reported no dispersion measure or confidence interval. The teleoperated tasks were qualitative demonstrations, and no task success rates were reported.
The authors identify operator-set stiffness, missing tactile information and possible extensions to quadrupedal or wheeled bases as future work. They also point to autonomous stiffness adjustment as a future direction.
The report is an arXiv preprint, version 1, dated 26 Aug 2026.
Paper data and sources
Original title: LAC: Linear and Angular Compliance for Humanoid Whole-body Control
Authors: Yang Liu, Zhongkai Gu, Wei Zhu, Mitsuhiro Hayashibe
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text