GOLEM: Modular Humanoid Autonomy Towards Electric Vehicle Battery Disassembly

Anonymous Authors

Under double-blind review

Video

Abstract

Disassembling end-of-life electric vehicle (EV) battery packs is dull and dangerous work, performed almost entirely by humans. We present GOLEM (Generalized Open Library of Embodied Modules), an end-to-end, open-source system architecture for EV battery disassembly with the Unitree H1-2 humanoid robot in which walking, manipulation, dynamic stability, navigation, and spatial memory are independent modules with abstract interfaces, so that methods are easily developed, interchanged, and compared. GOLEM is deployed as a Docker-based ROS 2 abstraction in which MuJoCo and IsaacLab digital twins expose interfaces matching the physical robot. GOLEM's composability and per-module customization enable development and demonstration of humanoid EV battery disassembly, from simulation to reality.

GOLEM provides fair comparison between humanoid modules, enabling evaluation as a capability ladder, in which one module is characterized at a time and added as a rung: LiDAR-inertial navigation places the robot within 13.0 cm of a 6 m goal; a learned standing controller recovers from external disturbances that sampling-based lower-body MPC does not; and grasping loosened fasteners from a real Hyundai Ioniq 5 pack degrades from 97% tethered to 87% free-standing to 37% under navigation-induced pose variance.

A Unitree H1-2 humanoid standing at a Hyundai Ioniq 5 battery pack beside a gantry-mounted robot arm.
A Unitree H1-2 humanoid working alongside a gantry-based disassembly cell on a Hyundai Ioniq 5 battery pack. GOLEM augments automation infrastructure designed around human workers with humanoid autonomy.

System

GOLEM system diagram: physics engines, hardware interfaces, capabilities, and skills.
Four interoperable, modular layers. GOLEM connects (1) three physics engines — the real world, MuJoCo, and PhysX via IsaacLab — with (2) hardware interfaces for motor control, proprioception, and exteroception, (3) core capabilities spanning learning- and planning-based bodily control, navigation, and semantic object localization, and (4) skills for contact-rich, precise manipulation and retrieval with spatial memory. Because the twins expose the same ROS 2/DDS interfaces as the physical robot, switching between simulation and hardware amounts to selecting a DDS domain.

Results

13.0 cm
mean endpoint error
over a 6 m walk-to-goal
93%
balance recovery under
push, learned vs. MPC
97 → 87 → 37%
real screw grasping across
three autonomy levels
Scatter plot of navigation endpoint errors for a 6 metre walk-to-goal task.
Navigation. Six-meter start-to-goal walking accuracy from noisy initialization, using a LiDAR-inertial odometry and Nav2 stack with an off-the-shelf walking policy. This endpoint error defines the base-placement variance that downstream manipulation must tolerate.
Push force and center-of-mass margin traces comparing MJPC and ALMI standing controllers.
Standing under external force: MPC vs. RL. (a) An instrumented rod with a six-axis load cell applies an oppositional force to the torso; (b–c) balance is lost when the center-of-mass margin to the support polygon goes negative. Both controllers show similar initial force adaptivity, but (b) sampling-based MPC cannot recover and must be emergency-stopped, while (c) adversarial RL recovers a positive margin.
Box and scatter plot of center-of-mass margin across grasp synthesis methods, tasks, and autonomy levels.
Stability is sensitive to the grasp synthesis method. Episode-level center-of-mass margin — the signed distance from the floor-projected center-of-mass to the support polygon boundary — across simulated fridge opening and simulated and real battery screw picking. Negative values indicate a recovery step or a fall.