EXOKERN1's picture
Add V2 temporal model results to dataset card
b91a40f verified
---
pretty_name: "EXOKERN ContactBench v0 — Peg Insertion with Force/Torque"
license: cc-by-nc-4.0
tags:
- robotics
- force-torque
- contact-rich
- manipulation
- insertion
- lerobot
- isaac-lab
- forge
- franka
- simulation
- benchmark
- physical-ai
- wrench
language:
- en
size_categories:
- 1K<n<10K
task_categories:
- robotics
dataset_info:
features:
- name: observation.state
dtype: float32
shape:
- 24
- name: observation.wrench
dtype: float32
shape:
- 6
- name: action
dtype: float32
shape:
- 7
- name: timestamp
dtype: float64
- name: frame_index
dtype: int64
- name: episode_index
dtype: int64
- name: index
dtype: int64
- name: task_index
dtype: int64
splits:
- name: train
num_examples: 330929
configs:
- config_name: default
data_files:
- split: train
path: "data/**/*.parquet"
---
# EXOKERN ContactBench v0 — Peg Insertion with Force/Torque
<p align="center">
<strong>The first publicly available insertion dataset with calibrated 6-axis force/torque annotations.</strong><br>
<em>Part of the <a href="https://huggingface.co/EXOKERN">ContactBench</a> collection by EXOKERN — The Data Engine for Physical AI</em>
</p>
<p align="center">
<img src="force_profile_sample.png" alt="Force profile during peg insertion episode" width="720">
</p>
---
## Why This Dataset Exists
Over 95% of existing robotics manipulation datasets contain **no force/torque data**. Yet contact-rich tasks — insertion, threading, snap-fit assembly — fundamentally depend on haptic feedback for reliable execution. Vision alone cannot distinguish a jammed peg from a seated one.
This dataset provides **2,221 peg-in-hole insertion episodes** with full **6-axis wrench data at every timestep**, generated using the [FORGE](https://arxiv.org/abs/2408.04587) (Force-Guided Exploration) framework in NVIDIA Isaac Lab. Every frame captures what the robot *feels*, not just what it sees.
---
## Dataset Overview
| Metric | Value |
|---|---|
| Episodes | 2,221 |
| Total Frames | 330,929 |
| Avg Episode Length | ~149 steps |
| Control Frequency | 20 Hz |
| Format | [LeRobot v3.0](https://github.com/huggingface/lerobot) (Parquet) |
| Robot | Franka Emika Panda (7-DOF) |
| Simulator | NVIDIA Isaac Lab 2.3.x + PhysX GPU |
| Task | `Isaac-Forge-PegInsert-Direct-v0` |
| Size | ~75 MB |
| License | CC-BY-NC 4.0 |
---
## Features
Each frame contains the following tensors:
| Feature | Shape | Description |
|---|---|---|
| `observation.state` | `(24,)` | Flattened observation vector — see [State Tensor Semantics](#state-tensor-semantics) below |
| **`observation.wrench`** | **`(6,)`** | **6-axis force/torque: [Fx, Fy, Fz, Mx, My, Mz] — see [Wrench Specification](#wrench-specification)** |
| `action` | `(7,)` | Delta end-effector pose command [dx, dy, dz, dRx, dRy, dRz] + success prediction |
| `timestamp` | scalar | Wall-clock time within episode (s) |
| `frame_index` | int | Frame position within episode |
| `episode_index` | int | Episode identifier |
---
## 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 | Field | Dims | Unit | Description |
|---|---|---|---|---|
| 0–2 | `fingertip_pos` | 3 | m | End-effector (fingertip) position in world frame |
| 3–5 | `fingertip_pos_rel_fixed` | 3 | m | EE position relative to the socket (fixed part) |
| 6–9 | `fingertip_quat` | 4 | — | EE orientation as quaternion [w, x, y, z] |
| 10–12 | `ee_linvel` | 3 | m/s | End-effector linear velocity |
| 13–15 | `ee_angvel` | 3 | rad/s | End-effector angular velocity |
| 16–21 | `force_sensor_smooth` | 6 | N, N·m | Smoothed 6-axis wrench (same data as `observation.wrench`) |
| 22 | `force_threshold` | 1 | N | Maximum allowable contact force (FORGE parameter) |
| 23 | `ema_factor` | 1 | — | Exponential moving average smoothing coefficient |
> **Implementation reference:** These fields correspond to `OBS_DIM_CFG` in [`factory_env_cfg.py`](https://github.com/isaac-sim/IsaacLab/blob/main/source/isaaclab_tasks/isaaclab_tasks/direct/factory/factory_env_cfg.py) (indices 0–15) extended by the FORGE force-sensing additions (indices 16–23). See [Noseworthy et al., 2024](https://arxiv.org/abs/2408.04587) §III for details.
---
## Wrench Specification
| Property | Value |
|---|---|
| **Coordinate frame** | **End-effector body frame** (Franka `panda_hand` link) |
| **Convention** | [Fx, Fy, Fz, Mx, My, Mz] |
| **Force unit** | Newtons (N) |
| **Torque unit** | Newton-meters (N·m) |
| **Source** | `env.unwrapped.force_sensor_smooth` — Isaac Lab contact sensor with EMA smoothing |
| **Typical force range** | ±50 N per axis |
| **Typical torque range** | ±10 N·m per axis |
The wrench is reported at the **end-effector body frame** attached to the Franka gripper link. 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).
> **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.
---
## Quick Start
```python
from lerobot.datasets.lerobot_dataset import LeRobotDataset
# Load dataset
dataset = LeRobotDataset("EXOKERN/contactbench-forge-peginsert-v0")
print(f"Episodes: {dataset.num_episodes}, Frames: {len(dataset)}")
# Access a single frame
frame = dataset[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")
# Decompose state vector
ee_pos = state[0:3] # fingertip position (m)
ee_pos_rel = state[3:6] # position relative to socket (m)
ee_quat = state[6:10] # orientation quaternion
ee_linvel = state[10:13] # linear velocity (m/s)
ee_angvel = state[13:16] # angular velocity (rad/s)
wrench_state = state[16:22] # force/torque (N, N·m) — same as observation.wrench
f_threshold = state[22] # force threshold (N)
ema = state[23] # EMA smoothing factor
```
### Load with pure HuggingFace Datasets
```python
from datasets import load_dataset
ds = load_dataset("EXOKERN/contactbench-forge-peginsert-v0", split="train")
print(ds[0].keys())
```
---
## Experimental Results: The Value of Force/Torque
We trained a Behavior Cloning (BC) policy on this dataset to ablate the impact of Force/Torque data on the Peg-In-Hole task. Both policies achieve a 100% insertion success rate, but the difference in physical execution is massive:
| Condition | Success Rate | Avg Contact Force (N) |
|-----------|--------------|-----------------------|
| Kinematics Only | 100.0% | 6.4 N |
| **With Force/Torque** | **100.0%** | **0.1 N** |
**Conclusion:** Both policies solve the geometric task. But the F/T-aware policy performs the insertion softly like an expert, reducing contact forces by **98.4%**. In industrial applications, this is the difference between a successful assembly and a damaged part.
### V2: Temporal Model (10-Frame Window)
Upgrading to a 1D-CNN that processes the last 10 frames improves overall prediction quality by 4x, and yields a 9.6% F/T advantage in offline MSE:
| Model | Condition | Val MSE | Avg Force (N) |
|-------|-----------|---------|---------------|
| V1 MLP | With F/T | 0.475 | 0.1 N |
| V1 MLP | Without F/T | 0.520 | 6.4 N |
| **V2 Temporal CNN** | **With F/T** | **0.119** | **3.0 N** |
| V2 Temporal CNN | Without F/T | 0.132 | 4.7 N |
The temporal model dramatically improves both conditions, but especially helps the blind (no-F/T) policy compensate via position/velocity trends — confirming that **direct force feedback remains irreplaceable for gentle contact control**.
---
## Data Collection
| Parameter | Value |
|---|---|
| **RL Algorithm** | rl_games PPO (~200 epochs, reward ~352) |
| **Environment** | `Isaac-Forge-PegInsert-Direct-v0` |
| **Collection mode** | Single-env rollout (`num_envs=1`), deterministic policy |
| **Episode horizon** | Fixed 149 steps (no early termination during collection) |
| **Sensor bandwidth** | 20 Hz (matched to control frequency) |
| **Contact dynamics** | PhysX GPU solver, Isaac Lab default parameters |
| **Domain randomization** | FORGE defaults (controller gains, friction, mass, dead-zone) |
---
## Intended Use
**Research applications:**
- Training and benchmarking force-aware manipulation policies
- Sim-to-real transfer studies for contact-rich assembly
- Hybrid vision + force/torque policy architectures
- Evaluating the effect of F/T data on manipulation performance
**Not intended for:**
- Direct deployment on physical robots without sim-to-real calibration
- Safety-critical applications without additional validation
---
## Reproduction
Methodology documentation available upon request.
Visit [exokern.com](https://exokern.com) for details.
---
## Related Work
This dataset builds on the following research:
- **FORGE** — Noseworthy et al., "[Force-Guided Exploration for Robust Contact-Rich Manipulation under Uncertainty](https://arxiv.org/abs/2408.04587)" (ISRR 2024)
- **Factory** — Narang et al., "[Factory: Fast Contact for Robotic Assembly](https://arxiv.org/abs/2205.03532)" (RSS 2022)
- **IndustReal** — Tang et al., "[IndustReal: Transferring Contact-Rich Assembly Tasks from Simulation to Reality](https://arxiv.org/abs/2305.17110)" (RSS 2023)
- **LeRobot** — Cadene et al., "[LeRobot: State-of-the-art Machine Learning for Real-World Robotics](https://github.com/huggingface/lerobot)" (2024)
---
## License
**CC-BY-NC 4.0** — Free for research and non-commercial use.
Commercial licensing and custom Contact Skill Packs available from EXOKERN.
Visit [exokern.com](https://exokern.com) for enterprise inquiries.
---
## 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](https://exokern.com) · 🤗 [huggingface.co/EXOKERN](https://huggingface.co/EXOKERN)
---
## Citation
```bibtex
@dataset{exokern_contactbench_peginsert_v0,
title = {ContactBench v0: Peg Insertion with Force/Torque},
author = {{EXOKERN}},
year = {2026},
url = {https://huggingface.co/datasets/EXOKERN/contactbench-forge-peginsert-v0},
note = {2,221 episodes, 330K frames, 6-axis F/T, LeRobot v3.0 format}
}
```