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 CorrellarXiv 2026 · arXiv

This paper, recently published by researchers at the University of Colorado Boulder and the University of Notre Dame, explores the precarious reality of deploying full-size humanoid robots in industrial settings. Using the Unitree H1-2 as a testbed, the authors reveal a fundamental paradox: in the world of humanoids, the "emergency stop" designed to keep us safe can often be the most dangerous thing in the room.

TL;DR

Traditional industrial safety assumes that cutting power leads to a safe state, but for a 70kg humanoid, cutting power means a catastrophic 600-joule collapse. This paper details a runtime safety layer for the Unitree H1-2 and proposes "graduated force authority"—a framework that grants robots power in stages based on demonstrated competence and explicit human responsibility.

What Problem Does This Paper Address?

The authors target the hazardous task of disassembling electric vehicle (EV) batteries. These packs are heavy (450 kg), high-voltage (800 V), and prone to thermal runaway. While gantry robots handle some tasks, the facilities are designed for humans, making full-size humanoids an attractive, "drop-in" solution.

Fig. 1: A Unitree H1-2 humanoid working alongside the gantry-based RAPID disassembly cell [2] on a Hyundai Ioniq 5 battery pack.

However, humanoids present a unique safety challenge. Unlike a fixed-base robotic arm that can engage brakes when power is cut, a humanoid stays upright only through active, high-frequency control. If you hit the e-stop, the robot doesn't just stop; it falls. Because of geometric scaling laws, a 1.78m robot like the H1-2 carries roughly three times the fall energy of its smaller counterparts (like the Unitree G1), releasing about 600 Joules of energy into whatever—or whoever—is nearby.

Methods

The researchers implemented their safety mechanisms within GOLEM (Generalized Open Library of Embodied Modules), an open-source ROS 2 stack.

The core of their approach is a Runtime Safety Layer:

  • Independence: It runs as a separate low-level process, isolated from the complex (and potentially unpredictable) learned or model-based controllers.
  • Enforcement: It monitors joint positions, velocities, and torques, enforcing limits at 90% of the manufacturer's specifications.
  • Fault Detection: It triggers a stop if state estimates become stale, heartbeats are lost, or the physical Arduino-based e-stop is pressed.
Fig. 2: Skills command the lower- and upper-body con- trollers. Every joint command they produce passes through the safety layer, a separate process that is independent of the learned and model-based control modules. It enforces joint position, velocity, and torque limits at 90% of the vendor specification, treats a stale state estimate or a lost heartbeat as a fault, and polls a microcontroller connected to a physical e-stop. Losing contact with the microcontroller also stops the robot.

To test the system, they compared two lower-body controllers: MJPC (a model-predictive controller) and ALMI (a learned policy trained against disturbances). They subjected the robot to instrumented push trials using a rod equipped with a six-axis load cell to see how the controllers—and the safety layer—handled loss of balance.

Datasets

While not a traditional "big data" paper, the study relies on:

  • Push Trial Data: 15 trials per controller (MJPC vs. ALMI) measuring push force and Center of Mass (CoM) margin.
  • GOLEM Capability Ladder: A benchmark previously established by the authors that tracks success rates across different deployment stages (tethered, free-standing, navigating, etc.).

Key Insights and Ideas

  • The E-Stop Paradox: In a humanoid, a "Category 0" stop (immediate power removal) creates a new hazard. The authors argue that industrial standards (like ISO 13850) fail to account for robots with actively controlled stability.
  • Limits are Context-Dependent: A torque limit that is safe for delicate manipulation might be accidentally triggered by the high-impact "spikes" of a normal walking gait. There is no "one-size-fits-all" limit.
  • Borrowing from Child Development: The authors draw a fascinating parallel between robot safety and developmental psychology. They argue that current robot safety is "heteronomous" (obeying fixed rules from the outside). We should instead move toward "graduated authority," similar to how we grant teenagers a driver's license only after they demonstrate judgment and "object permanence."
  • The Responsibility Gap: If a learned controller fails in an unpredictable way, who is to blame? The authors suggest that force authority should only be granted when responsibility is explicitly assigned to a specific person or organization.

Key Results

The push trials highlighted the stark difference between controller types:

  • MJPC (Model-Based): The robot failed to recover from pushes. The operator had to press the e-stop in every single trial to prevent a crash, with an average reaction time of 4.68 seconds.
  • ALMI (Learned Policy): This controller was significantly more robust, recovering a positive balance margin in 14 out of 15 trials without any human intervention.
Fig. 3: (A) Pushes are applied with a rod instrumented with an ATI Mini45 six-axis load cell. (B) Push force (black) and margin of the center of mass (CoM) to the support polygon (green; below the horizontal red line the CoM has left the polygon), mean and spread over 15 trials per controller. Left: under sampling-based MPC (MJPC), the margin decreases steadily after the push and the operator presses the e-stop (vertical red line) at 4.68±2.86 s. Right: the adversarially trained ALMI policy recovers and restores a positive margin in 14 of 15 trials without intervention. The safety layer and interfaces are identical in both conditions; only the lower-body controller differs.

Quantitatively, the H1-2's actuators (360 Nm at the knee) and its mass mean that a fall is not just a mechanical failure; it's a high-energy event that could trigger the very battery fires the robot is meant to prevent.

Limitations and Future Work

The authors are candid about the study's scope:

  • Narrow Evidence: Results are from a single robot in a controlled lab setting with a safety tether.
  • Implementation Gap: The proposed "stop hierarchy" (e.g., controlled descent vs. damping) is not yet fully implemented.
  • Perception Trust: Graduated authority relies on a perception stack to know when a human is nearby, but the safety layer was specifically designed not to trust complex upstream modules like perception.

Future work aims to develop verifiable balance fallbacks (like Control Barrier Functions) that are small enough to be formally analyzed, ensuring that even if the "smart" controller fails, the robot can at least kneel or sit down safely rather than collapsing.

Summary updated

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.
BibTeX
@misc{conway2026runtime,
  author = {Max Conway and Yutong Zhang and Allen Devaraj and William Xie and Alessandro Roncone and Nikolaus Correll},
  title = {{Runtime Safety and Emergency Stops on a Full-Size Humanoid: Deployment Lessons and a Case for Graduated Force Authority}},
  year = {2026},
  eprint = {2611.15928},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15928},
}
Figure