← Paper Digest

Runtime Safety and Emergency Stops on a Full-Size Humanoid: Deployment Lessons and a Case for Graduated Force Authority

Max Conway, Yutong Zhang, Allen Devaraj, William Xie, Alessandro Roncone, Nikolaus Correll

November 2026

TL;DR A runtime safety layer and hard-wired e-stop for an open-source Unitree H1-2 humanoid stack, tested by pushing the standing robot until it lost balance. Under sampling-based MPC the operator had to press the e-stop in every trial, which drops the 70 kg robot, while a learned policy recovered in 14 of 15.

What they do

Golem is an open-source ROS 2 autonomy stack for the 1.78 m, 70 kg Unitree H1-2, built for disassembling electric-vehicle battery packs. This extended abstract covers its safety parts: a runtime safety layer, a hard-wired e-stop, push trials in which the e-stop was needed, and an argument that robots should be granted force authority in stages, with a named responsible party.

How it works

  • Separate safety layer. Every joint command passes through a low-level process, independent of the control modules, that checks measured joint positions, velocities and torques against limits set at 90% of the manufacturer’s maximums. It stops the robot on any violation or when the state estimate goes stale.
  • E-stop. The layer polls an Arduino wired to a physical button; losing contact with the Arduino also stops the robot. The stop switches motor torque off (an IEC Category 0 stop).
  • Push trials. A rod with a load cell pushes the standing robot slightly above the torso center until its center of mass leaves the support polygon. Two lower-body controllers run behind the same safety layer: sampling-based MuJoCo-MPC (MJPC) and ALMI, a learned policy trained adversarially against upper-body disturbances.

Results

  • Peak push forces were similar (MJPC 20.4 ± 7.0 N, ALMI 22.1 ± 6.9 N). Under MJPC the margin fell to -0.166 m with a 16.3° peak lean, and the operator pressed the e-stop in every trial. ALMI reached -0.044 m and 6.8° and recovered in 14 of 15 trials.
  • Cutting torque drops the robot from a center-of-mass height of about 0.9 m, releasing roughly 600 J. A safety line caught every fall.
  • The limits misfired: the vendor’s walking policy exceeds the limits in its own released URDF, and an elbow-torque cap meant to limit pushing was exceeded by footstrike spikes during walking, which would cut power to the legs mid-stride.

Limitations

One robot, 15 standing push trials per controller, and no measurement of what a fall damages. The 90% margin and the elbow cap were set by convention rather than hazard analysis. None of the proposed fixes (a stop hierarchy, a verifiable balance fallback, graduated force envelopes) is implemented, and the force-authority section is an argument rather than an experiment.

Figures

From Figure 1: A Unitree H1-2 humanoid working alongside the gantry-based RAPID disassembly cell on a Hyundai Ioniq 5...
From Figure 1: A Unitree H1-2 humanoid working alongside the gantry-based RAPID disassembly cell on a Hyundai Ioniq 5 battery pack.
From Figure 3: Pushes are applied with a rod carrying an ATI Mini45 six-axis load cell. Push force (black) and center...
From Figure 3: Pushes are applied with a rod carrying an ATI Mini45 six-axis load cell. Push force (black) and center-of-mass margin to the support polygon (green), over 15 trials per controller. Under MJPC the margin keeps falling and the operator presses the e-stop at 4.68 ± 2.86 s; the learned ALMI policy restores a positive margin in 14 of 15 trials.
Abstract

Full-size humanoid robots are being considered for hazardous industrial work such as disassembling electric-vehicle batteries. We describe the safety mechanisms of an open-source autonomy stack for the Unitree H1-2 humanoid and what deploying them taught us. All joint commands pass through a safety layer that runs separately from the control software. It enforces joint position, velocity, and torque limits, and it stops the robot when a state estimate becomes stale or when the hard-wired emergency stop (e-stop) is pressed or disconnected. We pushed the standing robot until it lost balance. With a model-predictive lower-body controller it never recovered, and the operator had to press the e-stop in every trial. Pressing it removes balance control, and the robot falls. The safety layer's own limits trigger the same stop, and limits chosen for manipulation were exceeded during normal walking. Near an operator or a battery pack, the fall can be more dangerous than the cause of the stop. Joint limits bound how much force a robot can apply, but not when applying it is appropriate. Drawing on research on responsibility and on child and adolescent development, we therefore make a case for granting robots force authority in stages, with a person or organization explicitly responsible for its use.

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.