# Can an Open-Source Humanoid Stack Actually Disassemble EV Batteries?

A 17-author University of Colorado team has demonstrated end-to-end autonomous EV battery disassembly on a [Unitree Robotics](https://humanoidintel.ai/companies/unitree-robotics) H1-2 humanoid, and the performance numbers tell an unvarnished story: fastener grasp success collapses from 97% in tethered conditions to 37% once full autonomous navigation is chained into the pipeline. That single statistic captures both how far the field has come and exactly where the hard problems remain.

The system — GOLEM, short for Generalized Open Library of Embodied Modules — is an end-to-end, open-source architecture that treats walking, [loco-manipulation](https://humanoidintel.ai/glossary/loco-manipulation), dynamic stability, navigation, and spatial memory as independent modules with abstract interfaces. It runs as a Docker-based ROS 2 abstraction, with MuJoCo and IsaacLab digital twins that expose interfaces matching the physical robot, enabling [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) development without hardware access.

The target hardware is a real Hyundai Ioniq 5 battery pack — not a mock-up — which immediately differentiates this from most academic manipulation benchmarks. The authors quantify capability with a "ladder" methodology: characterize one module at a time before stacking it onto the next, isolating where performance degrades. LiDAR-inertial navigation places the robot within 13.0cm of a 6-meter goal. A learned standing controller recovers from external disturbances that a sampling-based lower-body MPC cannot handle. And fastener grasping degrades from 97% tethered → 87% free-standing → 37% under navigation-induced pose variance.

That 60-point gap between tethered and fully autonomous performance is the most important engineering signal in this paper.

---

## Why EV Battery Disassembly Is the Right Benchmark

EV battery disassembly is among the most economically urgent and physically hazardous tasks in the clean energy supply chain. Spent packs contain residual charge, toxic electrolytes, and structurally complex fastener arrays designed for assembly, not removal. Human workers perform this work almost entirely by hand today, according to the paper — a supply chain reality that will become acute as the first wave of mass-market EVs reaches end-of-life.

From a robotics difficulty standpoint, the task bundles nearly every capability that makes humanoids hard to deploy: bipedal navigation to an unstructured workstation, whole-body stabilization under manipulation loads, and [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) of small fasteners in constrained geometry. Choosing this as a benchmark rather than, say, box-moving or cable-routing reflects intellectual seriousness about what humanoids actually need to do.

The Hyundai Ioniq 5 pack is a real production component, not a custom test fixture. That grounding matters when evaluating transferability claims.

---

## The Modular Architecture: What GOLEM Actually Provides

GOLEM's core engineering contribution is the decoupling of capability domains behind abstract interfaces. This is architecturally significant for the community, not just for this specific task:

**Navigation:** LiDAR-inertial odometry achieves 13.0cm positioning error over a 6-meter traverse. The paper characterizes this independently before coupling it to manipulation — a methodological discipline that most demo-centric papers skip.

**Dynamic stability:** A learned standing controller outperforms a sampling-based lower-body MPC on disturbance recovery. This is consistent with the broader industry trend toward learned whole-body controllers, but the direct comparison against a classical baseline in the same experimental rig is valuable data.

**Grasping:** The 97% → 87% → 37% cascade across tethered, free-standing, and navigation-coupled conditions is the paper's most citable finding. It directly quantifies the compounding cost of integration — each additional degree of freedom in the deployment chain introduces error that manipulation modules were not trained to handle.

**Simulation infrastructure:** Docker + ROS 2 with MuJoCo and IsaacLab twins exposing matching interfaces lowers the barrier for research groups to develop and benchmark against GOLEM modules without physical H1-2 access. This is the kind of infrastructure the open-source humanoid community has been missing.

The source code is publicly available, which means this is immediately forkable. That's not a trivial contribution in a field where most institutional robotics research releases papers without runnable code.

---

## What the 37% Grasp Rate Actually Means

The number needs context before anyone dismisses or overhypes it. Thirty-seven percent under full navigation coupling means the system is completing autonomous end-to-end disassembly attempts — something most humanoid systems cannot do at all in unstructured settings outside a lab. It is not a deployment-ready number. It is a research baseline.

The authors frame the degradation as navigation-induced pose variance, which points to a specific engineering solution: tighter localization, better whole-body compensation for residual positioning error, or manipulation policies trained with navigation-equivalent pose distributions. The modular architecture makes all three approaches tractable to pursue independently — which is precisely the point of building GOLEM this way.

From an industry strategy perspective, the capability ladder methodology the team introduces deserves to be replicated. The humanoid field has a persistent problem with cherry-picked demos that obscure where integration costs accumulate. Evaluating each module independently and then measuring degradation under integration gives other teams a reproducible benchmark structure.

---

## Industry Implications

Several threads converge here that matter beyond this single paper:

**Open-source infrastructure as a forcing function.** The robotics community currently lacks shared benchmarking infrastructure for humanoid [whole-body control](https://humanoidintel.ai/glossary/whole-body-control). GOLEM's Docker/ROS 2/MuJoCo/IsaacLab stack, if adopted, gives research groups a common substrate to compare against. This is how manipulation benchmarks like DEXTERITY or LIBERO shaped the arm-robot research trajectory — open infrastructure pulls research effort toward comparable results.

**The sim-to-real gap is localized, not uniform.** GOLEM's per-module evaluation reveals that the sim-to-real gap is not evenly distributed across capability domains. Navigation accuracy looks tractable at 13.0cm error. Grasping in isolation looks strong at 97%. The gap concentrates at integration boundaries — pose variance propagating from navigation into manipulation. This suggests that cross-module robustness training, not individual module refinement, is the next bottleneck.

**EV disassembly as a commercial beachhead.** Industrial partners will read this paper and see a specific near-term use case with quantified capability gaps and a clear improvement roadmap. A 37% baseline on a real Ioniq 5 pack, with a documented path to improvement, is a more investable starting point than a 95% success rate on a custom test fixture that doesn't exist outside one lab.

**Unitree H1-2 as research platform.** The choice of the H1-2 is not incidental. Unitree's hardware is increasingly the Linux of humanoid research — affordable, available, and community-supported. GOLEM builds on that platform in a way that compounds its ecosystem value.

---

## Key Takeaways

- **GOLEM is a fully open-source, end-to-end humanoid system** for EV battery disassembly, built on Unitree H1-2, Docker, ROS 2, MuJoCo, and IsaacLab.
- **Fastener grasp success rates**: 97% tethered → 87% free-standing → 37% under navigation-induced pose variance on a real Hyundai Ioniq 5 battery pack.
- **LiDAR-inertial navigation** places the H1-2 within 13.0cm of a 6-meter goal.
- **A learned standing controller** outperforms sampling-based lower-body MPC on disturbance recovery.
- **The capability ladder methodology** — characterizing each module independently before integration — is a reproducible benchmarking contribution independent of the EV application.
- **The 60-point grasp degradation** under full autonomy pinpoints cross-module pose variance, not individual module failure, as the primary engineering challenge.
- **Source code is publicly available**, making this immediately forkable by the research community.

---

## Frequently Asked Questions

**What is GOLEM and what does it do?**
GOLEM (Generalized Open Library of Embodied Modules) is an open-source humanoid robot software architecture from researchers at the University of Colorado. It enables a Unitree H1-2 humanoid to autonomously navigate to, stabilize at, and disassemble fasteners from end-of-life electric vehicle battery packs. It structures walking, manipulation, navigation, and spatial memory as independent modules with abstract interfaces, making individual capabilities easy to swap and benchmark.

**What grasp success rate does GOLEM achieve on EV battery disassembly?**
The system achieves 97% grasp success when tethered (robot position fixed), 87% when free-standing, and 37% when operating under the full autonomous pipeline including LiDAR-inertial navigation. The primary cause of degradation is pose variance introduced by navigation positioning error.

**What humanoid robot does GOLEM run on?**
GOLEM is deployed and demonstrated on the Unitree H1-2 humanoid robot. The simulation infrastructure uses MuJoCo and IsaacLab digital twins that expose interfaces matching the physical platform.

**What simulation tools does GOLEM use for sim-to-real development?**
GOLEM uses a Docker-based ROS 2 abstraction in which both MuJoCo and IsaacLab digital twins expose interfaces that match the physical Unitree H1-2, allowing module development and testing without hardware access before deployment on the real robot.

**Why is EV battery disassembly a useful benchmark for humanoid robots?**
EV battery disassembly requires bipedal navigation to an unstructured workstation, whole-body stabilization under load, and dexterous manipulation of small fasteners — bundling the core difficulty axes of humanoid deployment into a single economically meaningful task. It also reflects a real labor shortage: the task is currently performed almost entirely by humans despite being hazardous work, making it a credible near-term commercial application.

**How does GOLEM's learned standing controller compare to classical methods?**
GOLEM's learned standing controller recovers from external disturbances that a sampling-based lower-body MPC fails to handle. The paper evaluates both approaches on the same hardware under the same conditions, providing a direct comparison rather than an isolated demo.