shashwatsaxena136/builderbench-scripted-demos

BuilderBench scripted-expert demonstrations

1,000 successful demonstrations (200 per task) for five tasks of BuilderBench (UR5e + Robotiq 2F-85 cube-building environment), generated with a scripted expert. BuilderBench itself ships no demonstrations; these were made for imitation-learning experiments (Qwen3.5 vs. Qwen3.5 + VGGT-Omega fused visual features).

Task	Name	Episodes	Expert success during collection
cube-1-task-1	place-cube	200	200 / 200
cube-2-task-1	stack (2)	200	197 / 200 attempts kept
cube-2-task-2	horizontal align	200	195 / 300
cube-3-task-1	stack (3)	200	192 / 220
cube-4-task-1	stack (4)	200	197 / 410

Only episodes that pass BuilderBench's native success check are included (every goal cube within 5 mm of its target, held for 25 control steps, effector back at the neutral pose when the task requires it).

How the data was generated

• Environment: BuilderBench CreativeCubeEnv at commit de9130b, native 5-D action interface (normalized [-1, 1] deltas: x, y, z in ±5 cm, yaw in ±0.3 rad, gripper opening in ±1), 50 Hz control, native success/termination/truncation. Changes, applied through a subclass:
  • two extra fixed cameras (left, right) next to the native front_pixels camera, all 512×512;
  • a uniform rigid xy offset of the whole initial cube layout in ±1 cm per episode (stock resets are identical);
  • visualize_info=False (no goal ghosts or success recoloring in the images); the goal is given as separate goal renders;
  • reproducible resets (native reset uses an unseeded action sampler for two stabilization steps; it is seeded here).
• Expert: BuilderBench's own pick_and_place motion primitive (CreativeCubeLanguageWrapper), placing goal cubes bottom-up (ascending goal z), followed by a closed-loop servo of the effector to the neutral pose, then zero actions until native termination. Every recorded action is the native normalized 5-D action actually executed.
• Rendering: MuJoCo EGL with the Mesa llvmpipe software renderer. Rendering with a different backend (e.g. NVIDIA EGL) produces small pixel differences (≈0.5/255 mean) to which some policies are sensitive; the stored simulator states let you re-render every step with your own renderer.

File format

One compressed NumPy archive per episode: demos/<task>/ep_seed<seed>.npz (the reset seed), plus demos/<task>/index.json with collection metadata and every attempt (kept or not).

Key	Shape	Description
actions	[T, 5] float32	normalized native action executed at each control step
proprio	[T, 7] float32	robot-only state before each step: effector xyz, cos/sin yaw, gripper opening, gripper contact
rewards, terminated, truncated, success	[T]	native per-step signals
sim_state	[T, D] float64	full MuJoCo state before each step: [time, qpos(nq), qvel(nv), act(na), ctrl(nu)] (enough to re-render or resume any step)
decision_steps	[K] int	control steps at which frames were rendered (every 4th step)
frames	[K, 3, 512, 512, 3] uint8	RGB views at decision_steps, camera order front_pixels, left, right, state before the action at that step
goal_frames	[3, 512, 512, 3] uint8	goal configuration rendered from the same cameras
goal_xyzs, goal_masks	[cubes, 3], [cubes]	goal positions and which cubes have a goal
diag_object_pose	[T, cubes, 7]	cube xyz + wxyz quaternion before each step (oracle state: diagnostics only, not a policy input)

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download("shashwatsaxena136/builderbench-scripted-demos",
                       "demos/cube-2-task-1/ep_seed000000.npz", repo_type="dataset")
ep = np.load(path)
print(ep["actions"].shape, ep["frames"].shape)

Policies trained on this data should evaluate on reset seeds outside the demo seeds (our evaluations use seeds ≥ 1,000,000). Code for the environment wrapper, expert, collection and training: bbfusion/env.py, bbfusion/expert.py, scripts/collect_demos.py.

Limitations

Clean scripted demonstrations from near-identical initial layouts (±1 cm), with no recovery behaviour; policies trained on them see few off-nominal states. Tasks were limited to those the scripted expert can solve.