EXOKERN/skill-forge-peginsert-v0

EXOKERN Skill v0 — Peg Insertion with Force/Torque

The first publicly available contact-rich manipulation skill trained with calibrated 6-axis force/torque data.

Part of the EXOKERN Skill Catalog — The Data Engine for Physical AI.

---

Overview

This model is a Diffusion Policy trained for precision peg insertion in simulation (NVIDIA Isaac Lab / FORGE). It demonstrates that force/torque (F/T) sensing data consistently reduces contact forces during assembly tasks.

Two model variants are provided:

• full_ft — Trained with joint states + 6-axis force/torque feedback (22-dim observation)
• no_ft — Trained with joint states only (16-dim observation)

---

Key Results

F/T Ablation (3 Seeds × 100 Episodes)

Condition	Success Rate	Avg Force (N)	Peak Force (N)
full_ft (with F/T)	100.0%	3.2 ± 0.5	10.5 ± 0.4
no_ft (without F/T)	100.0%	5.2 ± 0.1	12.1 ± 0.3
Force Reduction	—	38.2%	13.2%

Per-seed breakdown

Seed	full_ft Force	no_ft Force	Reduction
42	3.7 N	5.3 N	30.4%
123	3.4 N	5.0 N	32.3%
7	2.5 N	5.2 N	52.0%

F/T-aware policies produce 38% less contact force at identical success rates.

Cross-Architecture Comparison

This skill was also benchmarked against simpler architectures on the same dataset:

Architecture	full_ft Force	no_ft Force	F/T Reduction	Parameters
V1 (MLP, single-frame)	0.1 N	6.4 N	98.4%	534K
V2 (Temporal CNN, 10-frame)	3.0 N	4.7 N	36%	273K
Diffusion Policy (this model)	3.2 N	5.2 N	38.2%	71.3M

The F/T advantage is consistent and robust across all architectures and seeds.

---

Architecture

1D Temporal U-Net with FiLM conditioning (Diffusion Policy, Chi et al. 2023).

Component	Details
Observation encoder	MLP: obs_horizon × obs_dim → 256-dim condition vector
Noise prediction	1D U-Net with 3 encoder/decoder stages, channel mults (1, 2, 4)
Conditioning	FiLM (Feature-wise Linear Modulation) via timestep + observation
Noise schedule	Cosine schedule (100 training steps)
Sampling	DDIM, 16 denoising steps at inference
Action chunking	Predict 16 actions, execute 8 (receding horizon)
Parameters	71.3M (full_ft) / 71.3M (no_ft)
EMA	Exponential Moving Average, decay=0.995

Observation Space

Component	Dimensions	Included in
Joint positions	7	both
Joint velocities	7	both
Joint torques	2	both
Force/Torque (wrench)	6	full_ft only
Total per frame	22 / 16
Observation horizon	10 frames	both

Action Space

Component	Dimensions
Joint position targets	7
Prediction horizon	16 steps
Execution horizon	8 steps

---

Usage

Quick Start

import torch
from inference import DiffusionPolicyInference

# Load the full_ft model (with force/torque)
policy = DiffusionPolicyInference("full_ft_best_model.pt", device="cuda")

# During rollout:
obs = get_observation()  # (22,) for full_ft, (16,) for no_ft
# Feed observations over time to build the 10-frame window
policy.add_observation(obs)

# When ready to act:
actions = policy.get_actions()  # Returns up to 8 actions
for action in actions:
    env.step(action)

Loading Weights

checkpoint = torch.load("full_ft_best_model.pt", map_location="cuda")
print(f"Condition: {checkpoint['condition']}")
print(f"Val Loss:  {checkpoint['val_loss']:.6f}")
print(f"Obs Dim:   {checkpoint['obs_dim']}")

---

Training Details

Parameter	Value
Dataset	EXOKERN/contactbench-forge-peginsert-v0
Episodes	2,221 (330,929 frames)
Train/Val split	85% / 15% (episode-based)
Epochs	300
Batch size	256
Optimizer	AdamW (lr=1e-4, weight_decay=1e-4)
LR schedule	Cosine annealing (eta_min=1e-6)
Seeds	42, 123, 7
Training time	~3h per run (RTX 4090)
Normalization	Min-max to [-1, 1]

---

Evaluation Environment

Parameter	Value
Simulator	NVIDIA Isaac Lab (Isaac Sim 4.5)
Environment	Isaac-Forge-PegInsert-Direct-v0
Robot	Franka Emika Panda (7-DOF)
Control mode	Joint position targets
Physics step	120 Hz
Control frequency	15 Hz (decimation=8)
Episodes evaluated	100 per condition
Domain Randomization	None (v0)

---

Files

File	Description
full_ft_best_model.pt	Best checkpoint — trained WITH force/torque
no_ft_best_model.pt	Best checkpoint — trained WITHOUT force/torque
config.yaml	Model hyperparameters and training config
inference.py	Standalone inference script with action chunking
results.json	Full evaluation results (all seeds)
training_curve_full_ft.png	Training loss curves (full_ft)
training_curve_no_ft.png	Training loss curves (no_ft)

---

Limitations

• Sim-only: Trained and evaluated in simulation. No real-robot validation yet.
• Fixed environment: No domain randomization. Performance under physical variation is unknown.
• Single robot: Franka Emika Panda only. Joint-space actions, not task-space.
• Single task: Peg insertion only.

These are addressed in upcoming versions — see EXOKERN Roadmap.

---

Citation

@misc{exokern2026skill,
  title={EXOKERN Skill v0: Force/Torque-Aware Diffusion Policy for Peg Insertion},
  author={EXOKERN},
  year={2026},
  url={https://huggingface.co/EXOKERN/skill-forge-peginsert-v0},
}

---

Related

• Dataset: EXOKERN/contactbench-forge-peginsert-v0
• Organization: EXOKERN on HuggingFace
• Website: exokern.com
• Diffusion Policy Paper: Chi et al., 2023