EXOKERN ContactBench v0.1 — Peg Insertion with Domain Randomization + Force/Torque
Domain-randomized demonstration dataset for contact-rich peg insertion, collected with full 6-axis Force/Torque sensing. Successor to v0 with 2.3× more episodes, multi-layer domain randomization, and significantly wider force distributions for robust policy training.
Note: For stable downstream references and reproducible training, consider using the semantic-versioned v0.1.1 release.
Part of the ContactBench collection by EXOKERN.
Why v0.1 over v0?
v0 was collected under fixed, ideal conditions: single friction value, perfect alignment, nominal mass. The resulting policies achieve 100% success in simulation but may struggle under real-world variation.
v0.1 trains policies to handle the variation they will encounter in deployment: different surface finishes (friction 0.2–1.2), object mass uncertainty (±30%), gripper placement error (±8 mm, ±5 deg), and controller gain mismatch (±20–30%).
The F/T channel becomes more valuable under DR — our v0 ablation showed 38% average force reduction with F/T; we expect this gap to widen under domain randomization because force feedback is essential for adapting to varied contact conditions.
| Metric | v0 | v0.1 |
|---|---|---|
| Episodes | 2,221 | 5,000 |
| Frames | 330,929 | ~750,000 |
| Collection regime | Mostly fixed condition | Domain-randomized |
| Wrench signal | Yes (6-axis) | Yes (6-axis) |
| Format | LeRobot v3.0 | LeRobot v3.0 |
Dataset Statistics
| Metric | Value |
|---|---|
| Episodes | 5,000 |
| Total frames | ~750,000 |
| Avg episode length | 149 |
| Control Frequency | 20 Hz |
| Robot | Franka FR3 (7-DOF) |
| Simulator | NVIDIA Isaac Lab |
| Task | FORGE Peg Insert (Isaac-Forge-PegInsert-Direct-v0) |
| Format | LeRobot v3.0 (Parquet) |
Features
| Key | Shape | Description |
|---|---|---|
observation.state |
(24,) | Collapsed policy state tensor exported by the FORGE task |
observation.wrench |
(6,) | Force/Torque vector [Fx, Fy, Fz, Mx, My, Mz] in N / N·m |
action |
(7,) | Action vector used for rollout supervision |
timestamp |
scalar | Per-frame timestamp |
frame_index |
scalar | Frame index within episode |
episode_index |
scalar | Episode index |
index |
scalar | Global frame index |
task_index |
scalar | LeRobot task mapping index |
State Tensor Semantics
The observation.state tensor is a flat 24-element vector produced by FORGE's collapse_obs_dict(). It concatenates the following quantities in order:
| Index Range | Field | Dim | Unit |
|---|---|---|---|
| 0–2 | fingertip_pos |
3 | m |
| 3–5 | fingertip_pos_rel_fixed |
3 | m |
| 6–9 | fingertip_quat |
4 | — |
| 10–12 | ee_linvel |
3 | m/s |
| 13–15 | ee_angvel |
3 | rad/s |
| 16–21 | force_sensor_smooth |
6 | N / N·m |
| 22 | force_threshold |
1 | N |
| 23 | ema_factor |
1 | — |
Note: observation.wrench contains the same data as observation.state[16:22]. The wrench is stored as a dedicated column for direct access without index arithmetic.
Implementation reference: These fields correspond to OBS_DIM_CFG in factory_env_cfg.py (indices 0–15) extended by the FORGE force-sensing additions (indices 16–23). See Noseworthy et al., 2024 §III for details.
Wrench Specification
The wrench is reported at the end-effector body frame attached to the Franka gripper link (panda_hand). Forces are measured via Isaac Lab's get_link_incoming_joint_force() method applied to the wrist joint, then smoothed with an exponential moving average (EMA factor configurable, default 0.2).
| Component | Index | Unit | Description |
|---|---|---|---|
| Fx | 0 | N | Force along x-axis |
| Fy | 1 | N | Force along y-axis |
| Fz | 2 | N | Force along z-axis (insertion axis) |
| Mx | 3 | N·m | Torque about x-axis |
| My | 4 | N·m | Torque about y-axis |
| Mz | 5 | N·m | Torque about z-axis |
Domain Randomization Stack
All layers applied cumulatively during data generation (not post-processing):
| Layer | Parameter | Range | Application mode |
|---|---|---|---|
| 1 - Geometric init | Peg x/y offset | [-8 mm, +8 mm] |
per episode reset |
| 1 - Geometric init | Peg z offset | [-3 mm, +5 mm] |
per episode reset |
| 1 - Geometric init | Peg roll/pitch | ±5 deg |
per episode reset |
| 1 - Geometric init | Peg yaw | ±10 deg |
per episode reset |
| 2 - Object physics | Peg mass scale | [0.7, 1.3] |
reset/startup (runner dependent) |
| 3 - Contact materials | Static friction | [0.2, 1.2] |
per episode reset |
| 3 - Contact materials | Dynamic friction | [0.15, 0.9] |
per episode reset |
| 3 - Contact materials | Restitution | [0.0, 0.3] |
per episode reset |
| 4 - Robot dynamics | Stiffness scale | [0.8, 1.2] (log-uniform) |
per episode reset |
| 4 - Robot dynamics | Damping scale | [0.7, 1.3] (log-uniform) |
per episode reset |
| 6/7 - Noise models | Observation/action Gaussian noise | active in v0.1 run config | runtime attachment |
Implementation notes:
- Final Option-A production run used fixed joint profile
nominal(friction/armature ×1.0). - Dedicated joint friction/armature EventTerm exists in code but was disabled for this final production collection.
Not included in v0.1 (deferred to v1): Sensor noise injection, action noise, joint friction/armature randomization, geometry variation.
Force/Torque QC Snapshot
| Metric | Value |
|---|---|
| Episodes with wrench data | 5000 / 5000 (100%) |
| Episode length min/mean/max | 149 / 149.0 / 149 |
| Contact-like episodes (QC proxy) | 4298 / 5000 (86.0%) |
| Contact-like criterion | max(‖F_xyz‖) > 2.0 N per episode |
| Peak force quantiles (0/50/90/99/100%) | 0.04, 8.25, 11.98, 14.91, 24.26 N |
| Mean force quantiles (0/50/90/99/100%) | 0.02, 1.65, 5.69, 7.78, 9.14 N |
The wide force distribution (0.04–24.3 N peak) reflects the DR stack: low-friction + good alignment produces near-zero contact force, while high-friction + misalignment + heavy peg produces up to 24 N. This variation is intentional and critical for training robust policies.
The ~14% non-contact-like episodes are valid successful insertions where low friction and good alignment resulted in minimal contact forces — these are not failures.
Collection Infrastructure
- Environments: 1,024–2,048 parallel (NVIDIA Isaac Lab on RTX 4090)
- Policy: Scripted insertion policy with DR-aware success filtering
- Collection time: ~30 minutes wall-clock (massively parallelized)
- QC pipeline: Automated wrench validation, force distribution analysis, episode completeness checks
Quick Start
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("EXOKERN/contactbench-forge-peginsert-v0.1")
print(f"Episodes: {ds.num_episodes}, Frames: {len(ds)}")
frame = ds[0]
state = frame["observation.state"] # (24,) — full observation
wrench = frame["observation.wrench"] # (6,) — force/torque
# Decompose wrench
force = wrench[:3] # [Fx, Fy, Fz] in N
torque = wrench[3:] # [Mx, My, Mz] in N·m
print(f"Force: {force} N")
print(f"Torque: {torque} N·m")
Load with pure HuggingFace Datasets
from datasets import load_dataset
ds = load_dataset("EXOKERN/contactbench-forge-peginsert-v0.1", split="train")
print(ds[0].keys())
Intended Use
Designed for imitation learning and diffusion-policy training on contact-rich insertion, including:
- Force-aware policy learning (
full_ftvsno_ft) - Robustness studies under domain randomization
- Sim-to-real transfer preparation and stress-testing
Recommended training config (from v0 baseline):
- Architecture: Diffusion Policy (71.3M params)
- Seeds: 3 × 2 conditions (full_ft / no_ft) = 6 runs
- Epochs: 300, batch size 256, lr 1e-4
Not intended for direct safety-critical deployment without additional validation and real-system calibration.
Evaluation Tool
Use exokern-eval to benchmark trained policies against EXOKERN baselines:
pip install exokern-eval
exokern-eval --policy your_checkpoint.pt --env Isaac-Forge-PegInsert-Direct-v0 --episodes 100
Related Resources
| Resource | Link |
|---|---|
| v0 Dataset | EXOKERN/contactbench-forge-peginsert-v0 |
| v0.1.1 Dataset (production) | EXOKERN/contactbench-forge-peginsert-v0.1.1 |
| Skill v0 (trained model) | EXOKERN/skill-forge-peginsert-v0 |
| Skill v0.1.1 (trained model) | EXOKERN/skill-forge-peginsert-v0.1.1 |
| Eval CLI | github.com/Exokern/exokern_eval |
| EXOKERN Org | huggingface.co/EXOKERN |
Related Work
- FORGE — Noseworthy et al., "Force-Guided Exploration for Robust Contact-Rich Manipulation under Uncertainty" (ISRR 2024)
- Factory — Narang et al., "Factory: Fast Contact for Robotic Assembly" (RSS 2022)
- IndustReal — Tang et al., "IndustReal: Transferring Contact-Rich Assembly Tasks from Simulation to Reality" (RSS 2023)
- Diffusion Policy — Chi et al., "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion" (RSS 2023)
- LeRobot — Cadene et al., "LeRobot: State-of-the-art Machine Learning for Real-World Robotics" (2024)
Citation
@dataset{exokern_contactbench_peginsert_v01_2026,
title = {ContactBench Forge PegInsert v0.1: Domain-Randomized Manipulation with Force/Torque},
author = {{EXOKERN}},
year = {2026},
url = {https://huggingface.co/datasets/EXOKERN/contactbench-forge-peginsert-v0.1},
license = {CC-BY-NC-4.0}
}
License
CC-BY-NC 4.0 — Free for research and non-commercial use. For commercial licensing, contact EXOKERN.
About EXOKERN
EXOKERN — The Data Engine for Physical AI
We produce industrially calibrated force/torque manipulation data for enterprise robotics, humanoid manufacturers, and research institutions. Contact-rich. Sensor-annotated. Industrially validated.
🌐 exokern.com · 🤗 huggingface.co/EXOKERN · 🐙 github.com/Exokern
- Downloads last month
- 20