Datumwell Foam Tower
Public release / 公开数据
All files in this release are publicly downloadable, without payment or access approval. This public dataset is a sample of Datumwell’s data work. There is one repository for this release, with no separate sample/full access tier. For more data or a custom collection, contact allen@datumwell.io or discuss your requirements.
当前发布的所有文件均可直接公开下载，无需付款或访问审批。 当前公开数据集作为 Datumwell 的合作评估样本，不再区分 sample/full 访问版本。如需更多数据或定制采集，请直接联系 allen@datumwell.io，或在官网提交需求。
Public access does not change component licenses, attribution requirements or documented quality limits. 更多数据的范围、可用性与交付安排另行沟通；公开访问不改变各组件许可、归因义务及已记录的质量边界。
An independently specified, AA-inspired MuJoCo collection for “Make the tallest tower out of the foam blocks.” This repository contains 200 complete selected episodes, 699,853 rows, and 600 synchronized videos. The release covers 30 scene combinations. All episodes, videos and accompanying files are publicly accessible.
Dataset overview · Provenance and rights
Specifications
| Field | Specification |
|---|---|
| Scene objects | 3, 4, or 5; cuboid, cube, cylinder, triangular prism, cone |
| Release | N3: 60 · N4: 70 · N5: 70 |
| Release split | Train 140 · Validation 30 · Test 30 |
| Observations | Three RGB views: left, right, overhead · 384 × 288 · native 25 Hz |
| Video | H.264 · yuv420p · one rendered observation per saved native state |
| State / action | Seven measured joint positions / seven absolute position setpoints |
| Units | Five arm joints in radians · two synthetic jaw slides in meters |
| Dynamics | 1 kHz MuJoCo · 100 Hz body/contact trace · 1 kHz robot trace |
| Package | LeRobot v3 · Parquet · MP4 · scene/configuration and quality provenance |
Generation and task scope
A procedural controller grasps, lifts, transports, opens the jaws, and retreats. Blocks have free joints and advance through simulated contact and gravity, with zero equality or mocap constraints. The scene uses NexArm-derived kinematics and collision geometry with synthetic parallel fingers; exported arm imagery uses project-authored primitives.
The controller stacks upright shapes using a vertical tool reference. A finite search compares two base orders and, when needed, two lateral-placement variants per layout. Collection uses up to three layouts per controller revision; earlier controller-development attempts remain in the local audit history. Measured stable heights within 0.5 mm can be tied for support, travel, and order diversity. Per-episode tested maximum, selected height, and regret are retained.
Two-cone scenes use one apex-up cone in the upright tower and leave the other present on the table. This release includes 26 planned_subset_stack episodes; other positive targets are full_stack. Object presence and target masks distinguish an unused object from a padded slot. This release has 99 actual bottom-to-top shape-order signatures. Every multiset has at least two, except BEE, whose upright family uses cube→cone; both cone identities occur in this release.
Table geometry, block dimensions, masses, friction, contact parameters, cameras, seeds, and controller revisions are recorded. These are modeling assumptions: a 0.70 × 0.60 m elliptical tabletop, three size bands, density 80 kg/m³, block/table sliding friction 0.8, and pad friction 1.5. Foam is represented by rigid convex bodies with compliant contact. The scene and reference comparison are documented in scene and capability scope.
Splits and diversity
Full-scene shape multisets are disjoint across train, validation, and test. Layouts cover broad scatter, arcs, two clusters, staggered rows, and asymmetric L arrangements. Candidate and controller variants stay with their source quota group. Global normalization uses training episodes; image statistics use sampled native frames.
When a second cone is unused, active-tower subsets can overlap across splits: train scene BCDEE can build BCDE, while BCDE is also a test full scene. The split therefore measures full-scene multiset separation, rather than active-tower-subset separation. Symmetry-aware object assignment and phase-normalized actual TCP distances screen near-duplicate scenes and trajectories.
The 200 positives were selected from 556 completed production candidates. The retained local history contains 177 task/quality-excluded candidates, 179 eligible unselected candidates, and three interrupted legacy rollouts. Selection counts describe this collection process rather than a learned-policy success rate.
Quality and reproducibility
Each selected tower is robot-contact-free and stable for 201 samples spanning 2.0 s at 100 Hz, with position/height variation below 2 mm and angular variation below 2°. Final height is within 3 mm of the selected nominal upright target. Checks cover reachability, joint limits, actual arm velocity ≤0.8 rad/s and acceleration ≤8 rad/s², jaw velocity ≤0.15 m/s and acceleration ≤20 m/s², contact penetration ≤3 mm, and unexpected contacts. Contact pairs and robot dynamics are audited at 1 kHz.
All selected videos are fully decoded with native timestamps and sampled source-image correspondence checks. The official JavaScript LeRobot reader and Python LeRobot 0.4.3 / PyAV load every episode, with first, middle, and final image access checks. Seed-to-layout reconstruction covers all 200 scenes; same-runtime physical replay covers one selected episode from each of 30 multisets. Native-rate float32 action replay covers representative 3/4/5-object and unused-cone behavior. Declared contact-failure controls provide separate grasp/drop/collapse checks. See SUMMARY.json and provenance reports.
highest_any_object_top_m, peak_ground_connected_height_m, highest_sustained_stable_height_m, and final_stable_height_m distinguish held or transient height from sustained free-standing height. These are project simulation metrics. Official AA scoring, policy training outcomes, and hardware transfer are outside this release's validation.
Loading
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("allen-datumwell/datumwell-foam-tower", video_backend="pyav")
row = ds[0]
print(row["observation.state"].shape, row["action"].shape)
Download directly from this public repository; access approval and authentication are not required. The portable loader requires no robot assets. Optional physical replay scripts use separately authorized local assets through NEXARM_DESCRIPTION_DIR.
The default training adapter selects measured state and action supervision. Privileged object/contact quantities, teacher phases, planned order, terminal outcomes, and future states are retained as audit fields. Feature definitions and the training feature policy describe clocks, units, masks, and input selection.
Sources
Task inspiration: Artificial Analysis AA-Robotics. The package contains independently generated simulation data and authored scene/arm imagery. AA reference footage and official robot model files remain outside the package. The separately supplied Hiwonder/NexArm dependency and component provenance are described in RIGHTS.md.
- Downloads last month
- 546