EXOKERN/contactbench-forge-peginsert-v0.1.1

EXOKERN ContactBench v0.1.1 - Peg Insertion with Domain Randomization and Force/Torque

contactbench-forge-peginsert-v0.1.1 is the current semantic-versioned reference dataset in the EXOKERN ContactBench series. It packages 5,000 domain-randomized peg-insertion episodes with explicit 6-axis wrench signals for reproducible training and evaluation on the Hugging Face Hub.

Quick Facts

Item	Value
Episodes	5,000
Frames	745,000
Control frequency	20 Hz
Format	LeRobot v3.0-compatible parquet
Robot	Franka FR3
Simulator	NVIDIA Isaac Lab
Task	Isaac-Forge-PegInsert-Direct-v0
Access	Gated download on the Hugging Face Hub

Release Positioning

Release	Episodes	Frames	Role
v0	2,221	330,929	Fixed-condition baseline
v0.1	5,000	745,000	First domain-randomized release
v0.1.1	5,000	745,000	Stable reference release

Use v0.1.1 if you want a pinned dataset reference for training, evaluation, or citation.

Why This Dataset Exists

Contact-rich tasks do not reduce cleanly to geometry. Peg insertion, threading, and snap-fit assembly all depend on force regulation once contact begins. This dataset exists to make that signal explicit and to package it in a Hub-native format that works for reproducible policy training.

The paired model release for this dataset is EXOKERN/skill-forge-peginsert-v0.1.1.

Data Schema

Field	Shape	Description
observation.state	(24,)	Collapsed policy state vector exported by the collection task
observation.wrench	(6,)	End-effector wrench [Fx, Fy, Fz, Mx, My, Mz] in N / N*m
action	(7,)	Normalized controller action recorded during data collection
timestamp	scalar	Per-frame timestamp in seconds
frame_index	scalar	Frame index within an episode
episode_index	scalar	Episode identifier
index	scalar	Global frame index
task_index	scalar	LeRobot task mapping index

State Tensor Semantics

The observation.state vector is a 24D flat tensor exported by the FORGE collection pipeline. The current mapping is:

Index Range	Field	Dim	Unit
0:3	fingertip_pos	3	m
3:6	fingertip_pos_rel_fixed	3	m
6:10	fingertip_quat	4	unitless
10:13	ee_linvel	3	m/s
13:16	ee_angvel	3	rad/s
16:22	force_sensor_smooth	6	N / N*m
22	force_threshold	1	N
23	ema_factor	1	unitless

observation.wrench duplicates observation.state[16:22] as a dedicated column for direct access.

Wrench Convention

Property	Value
Frame	End-effector body frame attached to the gripper
Order	[Fx, Fy, Fz, Mx, My, Mz]
Force unit	Newtons
Torque unit	Newton-meters

Domain Randomization Stack

Layer	Parameter	Range	Mode
Geometric init	Peg x/y offset	[-8 mm, +8 mm]	Per reset
Geometric init	Peg z offset	[-3 mm, +5 mm]	Per reset
Geometric init	Peg roll/pitch	[-5 deg, +5 deg]	Per reset
Geometric init	Peg yaw	[-10 deg, +10 deg]	Per reset
Object physics	Peg mass scale	[0.7, 1.3]	Reset / startup
Contact materials	Static friction	[0.2, 1.2]	Per reset
Contact materials	Dynamic friction	[0.15, 0.9]	Per reset
Contact materials	Restitution	[0.0, 0.3]	Per reset
Robot dynamics	Stiffness scale	[0.8, 1.2]	Per reset
Robot dynamics	Damping scale	[0.7, 1.3]	Per reset
Runtime noise	Observation / action noise	Active in run config	Runtime

QC Snapshot

Metric	Value
Episodes with wrench data	5000 / 5000
Contact-like episodes	4298 / 5000
Contact-like threshold	`max(
Peak force quantiles (0/50/90/99/100%)	0.0388 / 8.2533 / 11.9820 / 14.9128 / 24.2634 N
Mean force quantiles (0/50/90/99/100%)	0.0189 / 1.6533 / 5.6933 / 7.7840 / 9.1356 N

This QC signal is a consistency check, not a success label.

Figure: representative wrench trace plus episode-wise mean and peak force coverage computed from the published parquet rollout data.

Recommended Pairing

The public paired model release is EXOKERN/skill-forge-peginsert-v0.1.1. In that repo, both full_ft and no_ft checkpoints reach 100% success across 600 total rollouts, while the force/torque advantage is inconclusive on this task under strong domain randomization.

That nuance is important: this dataset is useful both when a force/torque benefit appears and when robustness training closes the gap.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("EXOKERN/contactbench-forge-peginsert-v0.1.1")
print(ds.num_episodes, len(ds))

frame = ds[0]
state = frame["observation.state"]
wrench = frame["observation.wrench"]
action = frame["action"]

print(state.shape, wrench.shape, action.shape)

from datasets import load_dataset

ds = load_dataset("EXOKERN/contactbench-forge-peginsert-v0.1.1", split="train")
print(ds[0].keys())

Related Work

• FORGE: Force-Guided Exploration for Robust Contact-Rich Manipulation under Uncertainty
• Factory: Fast Contact for Robotic Assembly
• IndustReal: Transferring Contact-Rich Assembly Tasks from Simulation to Reality
• Diffusion Policy: Visuomotor Policy Learning via Action Diffusion

Citation

@dataset{exokern_contactbench_peginsert_v011_2026,
  title   = {EXOKERN ContactBench v0.1.1: Peg Insertion with Domain Randomization and Force/Torque},
  author  = {{EXOKERN}},
  year    = {2026},
  url     = {https://huggingface.co/datasets/EXOKERN/contactbench-forge-peginsert-v0.1.1},
  license = {CC-BY-NC-4.0}
}

Intended Use

• Robust imitation learning on contact-rich insertion
• Force-aware versus state-only policy comparisons
• Policy stress-testing under randomized contact conditions
• Reproducible training baselines on the Hugging Face Hub

Limitations

• Simulation only. The dataset is not a real-robot capture.
• The task scope is a single peg-insertion setup.
• Domain randomization ranges are documented here, but they do not span all real-world failure modes.
• Access is gated on the Hub.

Related Resources

• Paired model: EXOKERN/skill-forge-peginsert-v0.1.1
• Fixed-condition baseline: EXOKERN/contactbench-forge-peginsert-v0
• Organization page: huggingface.co/EXOKERN