EXOKERN/skill-forge-peginsert-v0

EXOKERN Skill v0 - Peg Insertion with Force/Torque

skill-forge-peginsert-v0 is the clean baseline skill release in the EXOKERN catalog. The repository contains a paired policy comparison on the fixed-condition EXOKERN ContactBench v0 dataset:

• full_ft_best_model.pt: primary checkpoint with 22D observations, including 6-axis force/torque input
• no_ft_best_model.pt: ablation baseline with the same architecture but without force/torque input

The main value of this release is not just success rate. It is a controlled reference point for measuring how force/torque sensing changes contact quality on a simple assembly task.

Quick Facts

Item	Value
Task	Peg insertion in simulation
Dataset	EXOKERN/contactbench-forge-peginsert-v0
Simulator	NVIDIA Isaac Lab (Isaac Sim 4.5)
Robot	Franka Emika Panda
Architecture	TemporalUNet1D diffusion policy
Parameters	71.3M
Observation horizon	10 frames
Prediction / execution horizon	16 / 8 actions
Primary checkpoint	full_ft_best_model.pt
Included ablation	no_ft_best_model.pt

Primary Benchmark

The Hub metadata for this repo tracks the primary full_ft checkpoint. The full repo includes the paired no_ft ablation for comparison.

Checkpoint	Success Rate	Avg Contact Force (N)	Peak Force (N)
full_ft	100.0	3.2 +/- 0.5	10.5 +/- 0.4
no_ft	100.0	5.2 +/- 0.1	12.1 +/- 0.3

Figure: published seed-wise benchmark summary for the fixed-condition baseline release.

Per-seed average contact force:

Seed	full_ft	no_ft	Reduction with F/T
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%

Takeaway: on this controlled baseline, force/torque input preserved success while substantially reducing contact force.

Architecture

This release uses a 1D Temporal U-Net diffusion policy with FiLM-style conditioning from the observation history and diffusion timestep.

Component	Value
Action dimension	7
Observation dimensions	22 (full_ft) / 16 (no_ft)
Diffusion training steps	100
DDIM inference steps	16
Base channels	256
Channel multipliers	(1, 2, 4)
Normalization	Min-max to [-1, 1]

Repository Contents

File	Description
full_ft_best_model.pt	Best checkpoint with force/torque input
no_ft_best_model.pt	Best checkpoint without force/torque input
inference.py	Self-contained inference helper and model definition
config.yaml	Training, dataset, and environment configuration
training_curve_full_ft_seed42.png	Training curve for full_ft, seed 42
training_curve_full_ft_seed123.png	Training curve for full_ft, seed 123
training_curve_full_ft_seed7.png	Training curve for full_ft, seed 7
training_curve_no_ft_seed42.png	Training curve for no_ft, seed 42
training_curve_no_ft_seed123.png	Training curve for no_ft, seed 123
training_curve_no_ft_seed7.png	Training curve for no_ft, seed 7

Usage

Reproduce evaluation with exokern-eval

pip install exokern-eval

wget https://huggingface.co/EXOKERN/skill-forge-peginsert-v0/resolve/main/full_ft_best_model.pt

exokern-eval \
  --policy full_ft_best_model.pt \
  --env Isaac-Forge-PegInsert-Direct-v0 \
  --episodes 100

Load the repo helper locally

import os
import sys

from huggingface_hub import snapshot_download

repo_dir = snapshot_download(
    repo_id="EXOKERN/skill-forge-peginsert-v0",
    allow_patterns=["*.pt", "inference.py"],
)
sys.path.insert(0, repo_dir)

from inference import DiffusionPolicyInference

policy = DiffusionPolicyInference(
    os.path.join(repo_dir, "full_ft_best_model.pt"),
    device="cpu",
)

policy.add_observation([0.0] * 22)
actions = policy.get_actions()
print(len(actions))

Training And Evaluation Setup

Item	Value
Train / val split	85% / 15% by episode
Epochs	300
Batch size	256
Optimizer	AdamW, lr=1e-4, weight_decay=1e-4
LR schedule	Cosine annealing to 1e-6
EMA decay	0.995
Seeds	42, 123, 7
Physics rate	120 Hz
Control rate	15 Hz
Domain randomization	Disabled in this release

Related Work

• FORGE: Force-Guided Exploration for Robust Contact-Rich Manipulation under Uncertainty
• Diffusion Policy: Visuomotor Policy Learning via Action Diffusion
• Factory: Fast Contact for Robotic Assembly

Citation

@misc{exokern_skill_peginsert_v0_2026,
  title        = {EXOKERN Skill v0: Peg Insertion with Force/Torque},
  author       = {{EXOKERN}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/EXOKERN/skill-forge-peginsert-v0}},
  note         = {Paired full_ft and no_ft diffusion-policy checkpoints}
}

Security Note

The checkpoints in this repo are PyTorch pickles. Load them only in a trusted or isolated environment after reviewing the repository contents.

Limitations

• Simulation only. This release does not claim real-robot readiness.
• The task is a relatively simple peg insertion setting with fixed conditions.
• Results should not be generalized to harder contact tasks without additional evidence.
• The repo exposes paired checkpoints for research comparison; the intended production-style reference in this repo is full_ft_best_model.pt.

Related Resources

• Dataset: EXOKERN/contactbench-forge-peginsert-v0
• Domain-randomized successor: EXOKERN/skill-forge-peginsert-v0.1.1
• Evaluation CLI: github.com/Exokern/exokern_eval
• Organization page: huggingface.co/EXOKERN