Datasets:

pcseanlo
/

RobotArena360_Assets

Per-scene digital-twin assets for evaluating vision-language-action (VLA) policies on 22 real DROID tabletop scenes reconstructed as interactable Gaussian-splat environments. Each scene ships the trimmed Gaussian splat, its collision mesh, the alignment between the reconstruction and the real robot frame, the fitted operation surface used for physics and shadow rendering, the reconstructed object meshes, and the original camera calibration, robot trajectory, and language instruction.

Total size is about 4.3 GB. Every file here is read by the evaluation code; nothing is intermediate or diagnostic output.

Layout

<scene>/
  interactable_background_assets/
    gaussian_trimmed.ply        # background Gaussian splat, floaters removed over the workspace
    collider.glb                # background collision mesh
  alignment/
    aligned_pose_refined.npy                 # splat/collider pose in the robot frame
    aligned_pose_scene_scale_refined.npy     # metric scale of the reconstruction
    operation_surface_plane_fit.json         # fitted table plane and footprint (physics)
    operation_surface_shadow_v2.json         # receiver descriptor (physics + lighting footprints)
    cropped_collision/                       # reachable-workspace crop of collider.glb + sidecar json
  metadata/
    intrinsics.npy, extrinsics.npy           # exterior camera calibration
    joints.npy                               # Franka joint trajectory of the source episode
    joints_old.npy                           # untrimmed source trajectory (droid2 only)
    lang.txt                                 # task instruction given to the policy
    masks/result.json                        # object segmentation index
    masks/transformations.json               # per-object pose, scale, and mesh key
    masks/physical_properties.json           # per-object mass / material hints
    physical_properties.json                 # root-level copy for scenes without the masks/ one
  assets/
    image_XXXX_segmented.glb                 # one reconstructed mesh per object in transformations.json
  scene_background/
    tabletop_hybrid/surface_texture.png      # table texture for the shadow-receiver render proxy
    (droid2 uses surface/surface_texture.png)
_global/
  task_mapping.json                          # scene -> instruction index (identical to each lang.txt)
  sysid/shared_sysid_v2.json                 # Franka joint gain profiles from system identification
  sysid/shared_sysid_v3_balanced.json

Scenes and tasks

Scene	Instruction
droid2	put toy bear into black bowl
droid4	move cup to the right of the table
droid7	pick up yellow cup
droid8	pick up the tape
droid17	Pick up the small bottle and put it on the left side of the table
droid26	Put the marker in the cup
droid28	Take the object and scrub the inside of the jug
droid34	Place the cup close to the sheet
droid35	Place the pen inside the yellow mug
droid37	Move the mug to the left
droid41	Move the grey tape to the right
droid45	Move the blue towel slightly to the right on the mat
droid69	stack cup on bowl
droid108	stack two cups on each other
droid138	put yellow block into blue cup
droid291	put yellow duck in pot
droid302	stack 2 bowls on top of one another
droid820	put lego into a bowl
droid950	Use cup to pour something granular
droid980	Stack 3 cups together
droid1034	Move the green brick into the bowl
droid1142	pour circular tokens into a bowl

Download

pip install -U huggingface_hub
hf download pcseanlo/RobotArena360_Assets --repo-type dataset --local-dir ./RobotArena360_assets

Or a single scene:

hf download pcseanlo/RobotArena360_Assets --repo-type dataset --include "droid820/*" "_global/*" --local-dir ./RobotArena360_assets

Conventions

Poses are stored as [x, y, z, qw, qx, qy, qz] in the robot base frame; quaternions are wxyz.
joints.npy is the 7-DoF Franka trajectory of the source DROID episode. Frame 0 is the evaluation reset pose under the recorded reset profile.
lang.txt is the instruction actually passed to the policy. _global/task_mapping.json is a convenience index with identical strings.
Object GLBs are keyed by the image_XXXX_segmented names in masks/transformations.json.

Source

Scenes are reconstructed from episodes of the DROID dataset. Object meshes and background splats were produced by the RobotArena360 scene generation pipeline; alignment and operation-surface fitting were produced by the evaluation repository that consumes these files.

Downloads last month: 143