Datumwell / Synthetic robot data
Foam Tower.
Built through
contact.
Prepared for AA-Robotics-style tower-building experiments: contact-based synthetic demonstrations with three native camera streams, joint targets and physical-state traces.
- 200
- complete episodes
- 25 Hz
- native images & actions
- 3–5
- blocks per scene
- 140 / 30 / 30
- train / validation / test
One complete episode · public index 6
Follow the grasp, release and retreat.
Complete native 25 fps recording at 1× speed. Switching views preserves playback time. Source: v4-validation-N4-ACDE-00-r3-l0-p0.
Representation & checks
The contact history stays attached.
Observations and commands
Three 384 × 288 RGB views, seven measured joint states and seven commanded joint-position targets at 25 Hz. 699,853 numeric rows and 600 complete videos in LeRobot v3.
Object poses, velocities, support contacts, pad forces, target masks and controller phases are retained as audit fields. The feature policy separates privileged state and planning labels from default policy inputs.
Physics and stability
MuJoCo advances at 1 kHz. Robot dynamics are audited at 1 kHz and object/contact state at 100 Hz. Blocks are free bodies, moved by jaw contact and gravity.
Every selected tower remains robot-contact-free over a final 2.0 s window, with 201 stability samples. Checks cover actual velocity, acceleration, joint limits, reachability, penetration and unexpected collision contacts.
Feature reference ↗ · Episode scorecard ↗ · Scene and capability boundary ↗
Scope and provenance
An inspectable simulation of a specific task.
Inspired by AA-Robotics’ foam-tower instruction. Table proportions, dimensions, mass, friction, camera calibration and scoring thresholds are recorded modeling assumptions. Shapes are rigid contact bodies; foam deformation is outside the model.
The planner searches a finite family of upright orders. A second cone may remain unused and is labeled separately. Recorded heights distinguish any-object peaks, ground-connected peaks, sustained stable height and the final free-standing tower. A global maximum height is not claimed.
Thirty full-scene shape multisets are separated across train, validation and test. Active tower subsets may overlap when an unused cone is present. Selected demonstrations are screened from 556 completed candidates; failures and eligible unselected candidates are cataloged separately. This selection fraction is not a learned-policy success rate.
Same-runtime seed replay covers all 30 multisets; representative exported-action replays and official Python/JavaScript LeRobot loading checks accompany the release. The package makes no claim of an official AA environment or score, real-robot validation, or verified training gains.
Robot visuals use project-authored primitives. A locally authorized NexArm-derived dependency supplies simulation kinematics and collision geometry; official model files and AA reference footage are excluded. Consult rights and usage terms ↗.
Inspect, then request access
Start with ten complete episodes.
The public subset contains six train, two validation and two test episodes across all five layout families. Source mappings preserve the link to the 200-episode collection. The complete 200-episode collection is ¥500 CNY, one-time, including the 10 free samples. We contact you to confirm payment and manually open access on Hugging Face within 8 hours of confirmed payment and receipt of your HF username and repository access request.