EXOKERN — Pre-trained Manipulation Skills for Physical AI

Force/torque-validated. Multi-seed evaluated. Open & reproducible.

We build pre-trained robotic manipulation skills for contact-rich tasks — where vision alone fails and precise force control matters. Every skill ships with full F/T ablation studies, multi-seed evaluation, and standardized benchmarks.

🎯 The Problem

Over 95% of robotic manipulation approaches are vision-only. They fail at contact-rich tasks like insertion, assembly, and snap-fit — where sub-Newton force control is the difference between success and broken parts.

Training a robot for a new contact task takes weeks of engineering. Validating that it works safely takes even longer.

💡 What EXOKERN Provides

Skills — Pre-trained policies ready for deployment

More skills in development: Screw Driving, Snap-Fit Assembly, Gear Meshing, Wire Routing, Pick-and-Place.

Datasets — Contact-rich training data with F/T annotations

Tools — Evaluate any manipulation policy

pip install exokern-eval
exokern-eval --policy your_checkpoint.pt --env Isaac-Forge-PegInsert-Direct-v0 --episodes 100

• PyPI: exokern-eval
• GitHub: github.com/Exokern/exokern_eval

📊 Validated Results — Peg Insert v0

Evaluated across 3 random seeds × 2 conditions × 100 episodes = 600 rollouts total:

• 100% success rate across all seeds and conditions
• 38% average force reduction with F/T-aware policies (3.2N vs 5.2N)
• Consistent across architectures: MLP, Temporal CNN, and Diffusion Policy all show F/T benefit
• Reproducible: All checkpoints, configs, and eval scripts published

🔧 What Makes EXOKERN Different

• Force/torque ablation on every skill — quantified proof that F/T data improves performance
• Multi-seed evaluation — not cherry-picked single runs, statistically validated results
• Open benchmarks — use exokern-eval to test your own policies against our baselines
• LeRobot v3.0 compatible — datasets plug directly into standard training pipelines
• Industrially relevant tasks — insertion, assembly, contact-rich manipulation
• Full data provenance — capture methodology, sensor specs, evaluation protocol documented

🏗️ Built With

• NVIDIA Isaac Lab for high-fidelity physics simulation
• Diffusion Policy architecture (71.3M parameters)
• Franka FR3 (7-DOF) robot platform
• LeRobot v3.0 format for interoperability
• Simulated 6-axis F/T sensing (Bota Systems SensONE modeled)

🗺️ Roadmap

Building toward a full catalog of contact-rich manipulation skills with multi-robot support:

See the full roadmap on GitHub.

📬 Contact

• 🤗 HuggingFace: huggingface.co/EXOKERN
• 🐙 GitHub: github.com/Exokern
• 📧 Contact: [email protected]

---

EXOKERN — Bridging the Haptic Gap in Robotic Manipulation