PaperlayerView exact revision on arXiv

Source-linked AI summary

GOLEM: Modular Humanoid Autonomy Towards Electric Vehicle Battery Disassembly

Max Conway, William Xie, Allen Devaraj, Yutong Zhang, Niraj Pudasaini, Mateo Feit, Adam Abid, Zachary Allen, Chen Liu, Xuan Tan, Jensen Lavering, Jason Chen, Lyle Antieau, Anthony Von Pischke, Alessandro Roncone, Zachary Sunberg, Nikolaus Correll

arXiv:2608.21550v1cs.RO

TL;DR

End-of-life EV battery disassembly is hazardous and largely manual. GOLEM provides an open-source, modular humanoid architecture with interchangeable interfaces across simulation and hardware, and demonstrates a capability ladder for real battery-pack screw removal. The evaluation reports 13.0cm navigation error over 6m, learned standing recovery from disturbances, and grasping degradation under increasing autonomy and pose variance.

  • Problem

    EV battery disassembly requires repetitive fastener removal from energy-dense packs that can combust, emit toxic fumes, or electrocute workers.

  • Method

    GOLEM is an open-source architecture that composes independent walking, manipulation, stability, and navigation modules through matching simulated and physical interfaces.

  • Results

    GOLEM demonstrates a capability-ladder evaluation from navigation and standing control to real Hyundai Ioniq 5 screw grasping across increasing autonomy levels.

  • Takeaways & Limitations

    GOLEM supports fair module comparisons and attributes screw-removal errors primarily to locomotion error while exposing simulation limits for contact-rich tasks.

Abstract

from arXiv · show

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 13.0cm of a 6m 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. Source code is available at the project page https://golem-humanoid.github.io

I. INTRODUCTION

GOLEM addresses hazardous, repetitive EV battery disassembly with a modular humanoid architecture designed for human-oriented facilities. Its capability-ladder evaluation demonstrates interchangeable modules from simulation through real battery-pack screw removal.

  • Motivation: End-of-life EV-LIB disassembly is repetitive and hazardous because energy-dense packs may combust, emit toxic fumes, or electrocute workers.Hundreds of fasteners must be removed to recover valuable minerals and metals.
  • Motivation: Dedicated automation lacks versatility across the large variety of battery-pack designs, motivating humanoids that can integrate into human-oriented facilities.The paper targets full-scale humanoids for the workspace and torque demands of a 400kg Hyundai Ioniq 5 pack.
  • System contribution: GOLEM makes walking, manipulation, dynamic stability, and navigation independent modules connected through abstract interfaces.The system includes MuJoCo and IsaacLab digital twins that expose interfaces matching the physical robot.
  • Evaluation: GOLEM demonstrates humanoid fastener removal at tethered, standing, and post-navigation pose-variance autonomy levels.This capability ladder characterizes module interplay and where simulation fidelity fails to predict real-world performance.
  • System contribution: GOLEM enables fair comparison of interchangeable humanoid methods through matching simulated and real interfaces.The platform also supports comparisons of multiple Language-Conditioned Spatial Memory methods.

II. RELATED WORK

Prior EV-battery robotics addresses limited pack classes, task planning, or specialized disassembly infrastructure, while GOLEM combines humanoid autonomy with human-oriented facilities and modular software tools.

  • Robotic Disassembly of EV Batteries: EV-battery disassembly remains largely manual, while prior robotic work targets smaller hybrid packs or task planning.Material recovery motivates the task, but existing coverage is limited.
  • Robotic Disassembly of EV Batteries: RAPID addresses full-size 800V packs and unscrewing but requires factories designed around its infrastructure.Fastener operations account for about 75% of manual disassembly labor.
  • Robotic Disassembly of EV Batteries: GOLEM is presented as complementary to RAPID by enabling humanoid screw removal in facilities designed for human workers.The paper demonstrates this capability on a commercial EV pack.
  • Humanoid Control: Existing humanoid controllers span learned whole-body policies, motion tracking, adversarial locomotion coupling, and sim-to-real loco-manipulation.The cited systems target both full-size H1-2 and smaller G1 platforms.
  • Modular Robot Software: GOLEM reduces engineering labor by connecting specialized tools through shared abstract interfaces and an end-to-end architecture.These tools include OMPL, Nav2, Pinocchio, ROS 2, MuJoCo, and IsaacLab.

A. Container-based robot architecture

GOLEM uses Docker and ROS 2 containers to expose equivalent simulated and physical robot interfaces, while independent safety and motor-control layers govern execution.

  • Container architecture: GOLEM’s simulation container runs MuJoCo or IsaacLab digital twins that publish interfaces matching the physical robot.IsaacLab currently has verified support only for hardware interfaces.
  • Container architecture: A second ROS 2 container launches locomotion, inverse kinematics, perception, SLAM, navigation, safety monitoring, grasping, and exploration services.Skills are exposed as ROS 2 actions.
  • Simulation and hardware: Interface equivalence makes switching between simulation and hardware a matter of selecting the DDS domain.The paper emphasizes interface equivalence rather than dynamic fidelity because screw engagement is not faithfully simulated.
  • Safety: The independent safety controller monitors joint positions, velocities, and torques and immediately stops motion when configured limits are exceeded.The configured limits are 90% of Unitree’s specified maximum values.
  • Safety: An Arduino-connected physical emergency stop provides human-supervisor intervention, and communication loss is itself a stop condition.The same safety interfaces can be exercised in simulation before deployment.
  • Motor control: The motor layer commands desired joint positions and gravity-compensation torque to 500Hz PD servo loops, with zero desired velocity providing damping.An integral bias torque addresses steady-state errors from model inaccuracies and unmodeled loads.
  • Hardware interfaces: MAGPIE Hands provide force-controlled parallel grasping with built-in 3D perception and up to 32N grasp force.The grippers support contact detection and compliant off-center grasping.

1) Proprioception:

GOLEM estimates the H1-2 base state from IMU and joint measurements, using planted-foot constraints to update a random-walk Kalman filter without horizontal proprioceptive position observability.

  • Proprioception: The H1-2’s lower-body joints have torque limits of 360Nm for knees, 220Nm for hips and torso, and 75Nm for ankles.Upper-body shoulder, elbow, and wrist limits are 120, 120, and 30Nm, respectively.
  • Proprioception: GOLEM tracks base velocity with a random-walk Kalman filter using IMU orientation, angular rate, and joint encoder measurements.Ground-truth center-of-mass measurements are unavailable.
  • Proprioception: Planted-foot constraints provide the filter’s measurement, with each foot contact contributing a kinematic observation.The contact model uses the translational Jacobian of the contacting foot.
  • Proprioception: Foot measurements are fused with contact-weighted, outlier-gated noise, while IMU orientation is retained and horizontal position is integrated open-loop.Horizontal position is unobservable from proprioception in this setup.

2) Exteroception:

GOLEM separates lower-body control behind a common velocity interface while combining learned, sampling-based, and vendor controllers. Its sensing suite supports navigation, approach, and precise manipulation across distance ranges.

  • Exteroception: The exteroceptive suite combines LiDAR navigation, forward-facing RGB-D localization, and hand-mounted close-range cameras for manipulation.The sensors cover navigation, mid-range approach, and 7cm–50cm precise manipulation.
  • Lower-Body Control: The lower-body interface maps commanded base twists to the H1-2’s 12 leg joints.Input twists are clamped to each controller’s valid envelope, while zero twist implicitly handles standing.
  • Controller Comparison: GOLEM wraps MJPC, ALMI, and Unitree’s vendor policy behind one interface for fair controller comparison.MJPC plans at 67Hz over a 1s horizon, while ALMI runs at 50Hz.
  • Model-Based Control: MJPC optimizes contact-dynamics trajectories using predictive sampling over task residuals including height, velocity, orientation, foot placement, and effort.The controller warm-starts each receding-horizon step with the time-shifted plan.
  • Learning-Based Control: ALMI trains lower-body locomotion in a two-player zero-sum game against an adversarial upper-body policy.The formulation targets velocity tracking under disturbances such as recoil, payload swings, and momentum shifts.

2) Upper-Body Control:

GOLEM generates upper-body motion through a combination of differential inverse kinematics and sampling-based planning. The choice depends on motion range, obstruction, and collision-avoidance requirements.

  • Upper-Body Control: The upper-body interface accepts two SE(3) end-effector poses and outputs motor commands to achieve them.Motion generation combines differential inverse kinematics for short unobstructed motions with sampling-based planning for larger motions requiring collision avoidance.
  • Differential Inverse Kinematics: Differential inverse kinematics tracks task-space residuals by solving a quadratic program over weighted tasks and joint-limit constraints.The resulting joint velocity is integrated to produce upper-body commands.
  • Sampling-Based Planning: RRT-Connect generates collision-free upper-body paths in the 15-DOF joint space for longer-range motions in clutter.Sampled configurations are checked against joint limits, self-collision, and workspace constraints.

3) Object Localization:

GOLEM exposes object localization through a modular interface that accepts interchangeable detection models. RGB detections are applied to RGB-D point clouds for downstream skills.

  • Object Localization: The object-localization interface accepts models that produce bounding boxes or segmentation masks.It expects detected objects to be returned as point clouds from RGB-D cameras.
  • Model Choices: GOLEM uses Gemini Robotics for generic scenes and fine-tuned YOLO-World for niche battery components.The battery-component detector targets screws, nuts, and bus bars, whose RGB detections are applied to RGB-D point clouds.

4) Navigation:

GOLEM’s navigation stack combines LiDAR-inertial mapping, occupancy-grid planning, and local path tracking behind the lower-body control interface. Its higher-level skills and memory interfaces connect navigation to object retrieval and manipulation.

  • Navigation: FAST-LIO builds LiDAR-inertial maps from the Livox MID-360, while Nav2 plans over a costmap containing 2D occupancy and 3D collision information.The global planner uses A* with an admissible Euclidean heuristic and trades path length against obstacle clearance.
  • Navigation: The local controller tracks the planned path and emits velocity commands to GOLEM’s general lower-body controller interface.This preserves the separation between navigation and locomotion control.
  • Skills: Skills orchestrate GOLEM capabilities through defined interfaces to complete high-level objectives.This places navigation within a higher-level coordination layer rather than treating it as an isolated task.
  • Skills: The grasping module compares open-world grasp synthesis strategies within a standardized skill interface.Detected object point clouds can feed neural six-DoF or classical grasp-pose methods.
  • Spatiotemporal Memory: Spatiotemporal memory stores visual-language features in a persistent top-down semantic map for querying objects beyond the current field of view.A query returns candidate locations, relevance scores, and backend metadata for subsequent navigation and local manipulation.

IV. EXPERIMENTS

GOLEM structures evaluation as a capability ladder, measuring navigation, standing robustness, and manipulation under progressively greater autonomy. Navigation achieved 13.0 cm mean error over 6 m, while ALMI recovered from disturbances that MJPC could not.

  • IV. EXPERIMENTS: The capability ladder passes navigation-induced placement variance into downstream screw-grasping evaluation at three autonomy levels.The levels are tethered, standing, and standing with pose variance after navigation.
  • A. Walking Policy and Navigation Accuracy: 13.0 cm mean radial error placed the robot within 6 m of the navigation goal in 19 of 20 trials.The reported standard deviations were σx=6.83 cm and σy=11.43 cm, with 95% of runs reaching the goal.
  • B. Force-Adaptivity: MJPC and ALMI experienced similar initial force adaptivity, but MJPC could not recover after its CoM margin became negative.MJPC required an emergency stop after losing balance, whereas ALMI restored a positive CoM margin.
  • B. Force-Adaptivity: ALMI recovered a positive CoM margin in 93% of trials, while reaching a minimum margin of −0.044 m.ALMI exhibited 6.8±1.4° peak pitch-axis lean; subsequent experiments used ALMI for lower-body standing.

C. Manipulation

GOLEM evaluates manipulation across simulation and real battery-pack tasks while exposing grasping and memory methods through common interfaces. Visual servoing supports real screw picking, but navigation-induced pose variance and extraction geometry remain important failure sources.

  • Grasping: 100% centroid and 96.7% GGX-F success were achieved for simulated fridge-handle opening under tethered autonomy.Standing reduced success by 20.0% for centroid and 26.7% for GGX-F.
  • Grasping: 100% C-VS and 90.0% centroid success were achieved for simulated screw picking, while GGX and GGX-F failed completely at both tiers.The comparison attributes the GGX and GGX-F failures to imprecision in grasp generation.
  • Grasping: 96.7% tethered and 86.7% standing task success were achieved by visual servoing for real Ioniq 5 screw picking.Open-loop centroid grasping failed completely despite similar simulated and real CoM margins, because positional errors from depth segmentation and camera calibration were not robustly handled.
  • Grasping: 11/15 successful grasps occurred at three near navigation poses, while three off-nominal poses failed completely because they were out of reach.Median stability decreased from 137.2 mm at near poses to 124.8 mm at off-nominal poses.
  • Grasping: Five real C-VS failures resulted from off-axis screw extraction that caused jamming and slip after the initial grasp closed.The paper attributes these failures to nonlinear OMPL plans at joint limits and insufficient gripper force against friction.
  • D. Case Study: Language-Conditioned Spatial Memory: VLM re-ranking improved retrieval and temporal disambiguation but increased query latency by approximately 10^3×.SigLIP+FAISS had the lowest latency but limited temporal awareness, whereas VLMaps often prioritized outdated observations after object movement.
  • D. Case Study: Language-Conditioned Spatial Memory: GOLEM exposed memory backends through common interfaces, allowing backend selection according to spatiotemporal and runtime requirements.The benchmark used language-conditioned spatial and temporal retrieval across controlled RoboCasa object-relocation episodes.

V. CONCLUSION

GOLEM provides identical interfaces across simulated and physical resources, enabling modular development and fair comparison of humanoid autonomy components. The evaluation also shows that simulation and controlled-setting performance do not establish readiness for unsupervised operation around humans.

  • GOLEM exposes simulated and physical resources through identical ROS 2/DDS interfaces, enabling seamless mixing of digital twins and real hardware.
  • Fair module comparisons with GOLEM indicate that screw-removal errors reside primarily in locomotion error.The finding motivates future work on precise relocation and repositioning.
  • Simulation underpredicts perception error and base-motion coupling, making simulated success rates non-predictive for contact-rich, precise tasks.
  • Benchmark performance in simulation or controlled settings does not establish readiness for unsupervised operation around humans.The paper frames component-level failure attribution as a safety concern for deployment alongside people.
  • Table I compares memory modules through GOLEM’s common interface using average success over 20 RoboCasa episodes with 95% confidence intervals.
Loading 2608.21550v1…