Matt Strong

Matt Strong

CS PhD at Stanford University

← Paper Digest

ReACT: Reflection-Consistent Modulation and Cross-Load Consistency for Humanoid Velocity Tracking Under Asymmetric Loads

Xianghui Wang, Aoran Fu, Liyou Luo, Shanze Wang, Xiaoyu Shen, Dan Zhang, Wei Zhang

November 2026

TL;DR A humanoid carrying an object in one hand drifts sideways and in yaw. ReACT feeds an estimated load wrench into a mirror-symmetric policy and trains loaded and unloaded rollouts to move alike, and a Unitree G1 carrying 1.5 kg in one hand then reaches the goal collision-free in 40/40 navigation trials where the FALCON baseline fails all 40.

What they do

A humanoid that carries an object in one hand has a persistent asymmetric upper-body load. Walking stays stable, but velocity tracking picks up lateral and yaw-rate biases that accumulate into position and heading errors during navigation. ReACT is an RL locomotion controller for the Unitree G1 that keeps velocity tracking accurate, and nearly the same, whether the load is in the left hand, the right hand or absent.

How it works

  • Wrench estimate. A learned estimator predicts the 6D external wrench at the torso from a 15-frame proprioceptive history. It is trained on simulated labels and frozen during policy training.
  • Reflection-consistent modulation. Moving a load to the other hand keeps the vertical force and flips the roll moment. The actor splits its features into reflection-even and reflection-odd parts, and the wrench scales and couples them, so the policy compensates for the load while mirrored observations still give exactly mirrored actions.
  • Cross-load consistency. Matched triplets of environments (unloaded, left 4 kg, right 4 kg) share initial states and commands. A reward penalizes each branch’s displacement and heading change over a 0.4 s window for deviating from the triplet mean.
  • Training. FastSAC in Isaac Gym with 4096 environments and 0.5-6 kg hand payloads; the policy drives the 12 leg and 3 waist joints.

Results

  • Simulation, against the released FALCON checkpoint: forward, lateral and yaw-rate tracking MAE drop by 38.8-52.9%, 32.0-33.5% and 81.6-84.4%, and normalized torque cost by 38.9-40.9%.
  • Ablations: removing the consistency reward raises the lateral bias range across unloaded/left/right 4 kg at 0.8 m/s from 0.0114 to 0.2545 m/s; ReACT cuts yaw-rate MAE by 46.1-60.0% relative to that variant.
  • Real G1, mapless navigation with a fixed navigation policy and a 1.5 kg load in one hand: ReACT succeeds in 40/40 trials, FALCON collides in all 40.
  • Real-robot turning at 1 m/s: ReACT has lower forward (0.098 vs 0.202 m/s) and yaw-rate MAE, but higher lateral MAE (0.159 vs 0.135 m/s).

Limitations

Loads are static; changes of load during manipulation are left to future work. Real-robot tests cover one payload (1.5 kg), two scenarios and a narrow range of commands. Simulation results use two training seeds, FALCON is a single released checkpoint, and the concurrent ADAPT method is not compared quantitatively.

Figures

From Figure 1: Overview of ReACT. A wrench estimator feeds the disturbance modulation blocks that act on the symmetri...
From Figure 1: Overview of ReACT. A wrench estimator feeds the disturbance modulation blocks that act on the symmetric and antisymmetric actor features, and matched left-load, right-load and no-load rollouts are compared for the cross-load consistency reward.
From Figure 2: Signed lateral-velocity (a) and yaw-rate (b) bias versus unilateral payload mass at a 1.0 m/s forward ...
From Figure 2: Signed lateral-velocity (a) and yaw-rate (b) bias versus unilateral payload mass at a 1.0 m/s forward command, using an illustrative checkpoint of each method. Negative mass denotes left-hand loading, positive denotes right-hand loading.
From Figure 3: Planar trajectories for straight walking and quarter turns at 1.0 m/s. Top: 2 kg; bottom: 4 kg. Dashed...
From Figure 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.
From Figure 4: Real-world navigation comparison between ReACT and FALCON under unilateral hand loads of 1.5 kg. S and...
From Figure 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.
Abstract

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.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.