← Paper Digest

Leaning Over to Reach Far-Away Objects

Allen Devaraj, William Xie, Nikolaus Correll

November 2026

The H1-2 leans in, braces its forearm on the table, reaches across and recovers (2x speed).

TL;DR A sampling-based MPC planning 20 MuJoCo rollouts at 34 Hz on a laptop lets a Unitree H1-2 humanoid brace its forearm on a table and reach about 50 cm beyond its unbraced workspace. On hardware it reaches the commanded target in 89% of 54 trials and completes the full reach-and-recovery cycle in 81%.

What they do

People reach far across a table by leaning in and bracing a forearm on it. This paper gets a Unitree H1-2 to do the same with an online whole-body controller instead of a learned policy or a precomputed contact plan: it leans, braces, reaches a commanded point, holds, and recovers to standing, using onboard sensing and AprilTag fiducials (no motion capture), and transfers sim-to-real.

How it works

  • Sampling-based MPC in MuJoCo (MJPC). Every step draws 20 perturbed control sequences over a 1 s horizon, rolls them out in the MuJoCo model and averages the cross-entropy-method elites. It plans at a median 34 Hz on a laptop CPU.
  • Costs instead of a contact schedule. 39 residuals in nine groups: brace terms pull the chosen links onto the table, keep-out terms hold the other links off it, and reach, capture-point balance (over a support region that grows once the brace is seated), posture and regularization terms complete the task.
  • Sensing. IMU and leg odometry fused with AprilTag position fixes from the head camera; the wrist camera tracks a 5 cm target cube for visual servoing.

Results

  • Maximum reach from the ankle midpoint is 130±10 cm on hardware, about 50 cm beyond the ~80 cm unbraced reach; the braced posture holds for over four minutes.
  • Across three table heights and six targets, with no retuning: 89% (48/54) of reaches succeed and 81% (44/54) also recover, with 2.7±0.9 cm mean target error.
  • Fiducials cut horizontal estimator drift from +267 mm/min with proprioception alone to +4 mm/min.
  • Brace-hold success rises from 56% with six brace cost terms to 100% (15/15) with all nine. Under simulated model mismatch CEM completes 80/108 trials, against 42-62/108 for iCEM, predictive sampling and MPPI.

Limitations

The authors note that 81% full-cycle success is not yet reliable enough for autonomous deployment. Failures come from the forearm brace landing too far forward, which they attribute to sampling variance, and the deepest lateral targets on the tallest table are out of reach. The table needed rubber tape for friction, and absolute position relies on fiducial markers.

Figures

From Figure 1: Three stages of leaning and reaching on the real H1-2 (top) and in MuJoCo (bottom). The robot holds th...
From Figure 1: Three stages of leaning and reaching on the real H1-2 (top) and in MuJoCo (bottom). The robot holds the 1.3 m reach (frame 3), 50 cm past its unbraced standing reach of about 80 cm.
From Figure 7: The same controller at three table heights and six targets, three trials each. Dots are wrist-camera m...
From Figure 7: The same controller at three table heights and six targets, three trials each. Dots are wrist-camera measured hold positions, crosses the commanded positions, dotted circles a 4 cm radius, and x marks targets where the brace failed before reaching. At the tallest table the two deeper lateral targets fail.
From Figure 4: Reach, support margin, hand jitter above 1 Hz and bracing-arm force for commanded (left pair) and maxi...
From Figure 4: Reach, support margin, hand jitter above 1 Hz and bracing-arm force for commanded (left pair) and maximum-extent (right pair) reaches, in simulation (orange) and on hardware (green). Bars are medians over successful trials.
Abstract

We study supported reaching to extend a humanoid robot's workspace by bracing against the environment for stabilization using a straightforward yet capable whole-body model predictive control (MPC) approach, exclusively with local sensing and fiducial markers. While typical solutions for whole body humanoid control aim at avoiding or accommodating contact, contact is explicitly required to reach objects placed on surfaces far out of normal reach. We employ sampling-based MPC for whole-body leaning, bracing, reaching, and recovery on a Unitree H1-2 humanoid. The controller runs in real-time at 34 Hz on a laptop, transfers sim-to-real, extends reach by about 50 cm beyond the unbraced standing workspace, and sustains the loaded posture for over four minutes, before pose estimation drifts. The controller reaches the commanded target in 89% of trials and completes the full reach-and-recovery cycle in 81%, across 54 trials spanning six targets and three table heights. We further characterize why such a small real-time sampling budget is sufficient, finding that temporal persistence of sampled actions and exploration scale strongly govern success, while CEM provides the widest operating region over planning rate and model mismatch. These results establish a real-world baseline for online, contact-rich humanoid manipulation. Code, videos, and full hardware details are available at https://h12leaner.github.io/

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.