AssembleBench / README.md
lukasskellijs's picture
docs: update card for AssembleBench rename
016fefe verified
|
Raw
History Blame Contribute Delete
4.52 kB
metadata
license: apache-2.0
task_categories:
  - robotics
tags:
  - LeRobot
  - robotics
  - isaac-sim
  - assembly
  - manipulation
  - imitation-learning
  - contact-rich
configs:
  - config_name: default
    data_files: data/*/*.parquet

AssembleBench

Multi-task contact-rich assembly demonstrations for imitation learning / VLA fine-tuning, and the training set behind the AssembleBench benchmark: 14 tasks on the NIST ATB-1 taskboard (insert round and square pegs, mesh gears, thread nuts) on the DROID robot platform.

Read the writeup: Benchmarking Robot Models on Contact-Rich Assembly.

This dataset was created using LeRobot.

Summary

Episodes 1355
Frames 444651
FPS 15
Tasks 14 (pegs, gears, nuts; no M20)
Action DROID 8-D joint position (7 arm + binary gripper)
Cameras observation.images.front, observation.images.wrist at 180x320
Proprio observation.state (8), joint_vel (7), eef_pos (3), eef_quat (4)

Tasks (episode counts)

Language-conditioned pick / insert / mesh / thread:

family task n
peg_round 4 / 8 / 12 / 16 mm 96 / 99 / 100 / 100
peg_square 4 / 8 / 12 / 16 mm 98 / 98 / 99 / 100
gear_mesh small / medium / large 92 / 97 / 92
nut_thread M8 / M12 / M16 90 / 95 / 99

M20 removed: embedding analysis (PCA / t-SNE) showed M20 as a separate mode that split the nut cluster; it is not included.

How it was built

Built from a cleaned and topped-up revision of lukasskellijs/assembly_bench_2, with a second QC pass for duplicates, idle chunks, length outliers, and within-family Mahalanobis outliers. Pipeline:

  1. Start from the cleaned v2 revision (911 eps: dups + long outliers already removed from the original 1500).
  2. Drop all M20 nut episodes.
  3. Append scripted-expert top-up HDF5 demos that refill the gaps left by the clean pass (success-filtered Isaac Lab recordings).
  4. QC blocklist (45 trajectories by MD5 of full action || state):
    • exact-duplicate extras (none remained after clean + ingest dedup)
    • idle runs >= 15 frames (action-chunk size used in training)
    • per-task length >3 sigma and clear within-family 20-PC Mahalanobis outliers (retry / thrash / bad seat)
  5. Final exact-duplicate scan must be zero or the build aborts.

Quality over quantity: mild multivariate peg shape outliers that look like valid successes were kept.

Schema

DROID joint-position contract:

key shape notes
action (8,) absolute joint targets + gripper
observation.state (8,) joint pos (7) + gripper
observation.joint_vel (7,) rad/s
observation.eef_pos (3,) base-frame XYZ (m)
observation.eef_quat (4,) world WXYZ
observation.images.front video HxWx3 180x320
observation.images.wrist video HxWx3 180x320

Load

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("hud-evals/AssembleBench")
print(ds.num_episodes, ds.num_frames)
print(ds[0]["observation.images.wrist"].shape)

Intended use

  • Supervised fine-tuning of VLA / diffusion / ACT policies with action chunk size 15
  • Multi-task assembly benchmarking (peg / gear / nut families)

Not a teleop human dataset. Demonstrations are from a privileged scripted expert in Isaac Sim / Isaac Lab; visuals and proprio match the policy observation contract.

Reference checkpoints

pi0.5 finetunes trained on this dataset:

Checkpoint What it is
pi05-AssembleBench-12k behavior-cloning baseline, all 14 tasks
pi05-AssembleBench-cgdagger-r3 BC + 3 rounds of code-gated DAgger, the final checkpoint

Lineage

  • Simulator: NVIDIA Isaac Sim 6 / Isaac Lab Arena
  • Format: LeRobot v3.0