physicl/kitchen-workspace-understanding-safe-manipulation

Kitchen Workspace Understanding & Safe Manipulation

A kitchen rendered 400 times with relative depth, surface normals and per-material segmentation, for understanding counters and objects before a robot reaches for them.

_{12 of the 100 layouts from the same camera, then one view with its depth, segmentation and normals.}

At a glance

Scene	A kitchen with a marble island, bar stools and stocked counters.
Views	400 (100 layouts × 2 framings × 2 lighting setups)
Resolution	2048 × 1536
Format	PNG files, one row per view in a Parquet index
Layers per view	11 (image, depth, normals, material segmentation and more, listed below)
Camera	Two fixed framings per layout (43 mm, 60 mm)
Source	Rendered by Physicl from real 3D geometry with known materials

What changes from one view to the next

100 layouts of the same room, with the walls and the camera fixed:

• The four seats along the island change every layout: wooden bar stools, metal stools, upholstered stools, or rows of wheelie bins and slatted sorting bins in their place.
• Counters change too: cutting boards, kettles, flowers, a watering can, sometimes an ankle boot left on the worktop.
• Pendant lamps swap, and some layouts hang numbered signs from the ceiling.

Every layout comes in two lights (bright daylight and warmer, dimmer late-day light) and two framings: the wide view, plus a tighter second view of the island from the same camera (43 mm and 60 mm).

Within one framing, every view lines up with the others: same pixel, same place in the room.

_{Four of the 100 layouts, same camera.}

Ground truth on every view

Every layer comes from the 3D scene at render time. Nothing is estimated from the image afterwards, so they hold up in shadow, under strong sun and behind clutter.

Column	What it is	Format
image	Rendered image (beauty pass)	RGB PNG
depth_rel	Relative depth, rescaled per frame (not in metres)	8-bit PNG
colored_depth	Depth with a colour map, for viewing	RGB PNG
normal_world_gl_srgb	World-space surface normals, OpenGL convention, sRGB	RGB PNG
normal_world_gl_linear	Same, linear values	PNG
normal_world_dx_srgb	World-space surface normals, DirectX convention, sRGB	RGB PNG
normal_world_dx_linear	Same, linear values	PNG
material_index	One flat colour per material: a per-material segmentation (not per object instance)	RGB PNG
albedo	Surface colour without lighting	RGB PNG
attenuation	Light attenuation layer	PNG
uvw	Texture coordinates	PNG

Each layer column holds a URL to the PNG file in this repository, pinned to a fixed commit.

Load it

import io, requests, pandas as pd
from PIL import Image

df = pd.read_parquet("hf://datasets/physicl/kitchen-workspace-understanding-safe-manipulation/viewer/renders/train-00000-of-00001.parquet")
row = df.iloc[0]

image = Image.open(io.BytesIO(requests.get(row["image"]).content))
depth = Image.open(io.BytesIO(requests.get(row["depth_rel"]).content))
normals = Image.open(io.BytesIO(requests.get(row["normal_world_gl_srgb"]).content))

Other useful columns: file_name (layout number at the end, dl for late-day light, F1 or F2 for the second framing), camera_name, focal, light_id, combination_id (one id per layout).

What it is useful for

• Train counter and object understanding for kitchen manipulation: what sits where, and at what depth
• Learn object and surface geometry from rendered normals and relative depth
• Add clutter, object and lighting variety to kitchen perception data

About Physicl

Physicl builds the worlds physical AI learns from. This dataset is one of our early open releases. It was made before Physicl Studio and is not a Studio output.

Our product today is Physicl Studio: you describe the situation your model gets wrong and get it back as images or video, varied thousands of ways, with exact ground truth on every frame.

Studio is free for research teams during the beta: request access.

More open datasets: huggingface.co/physicl

Citation

@misc{physicl_kitchen_workspace_understanding_safe_manipulation,
  title  = {Kitchen Workspace Understanding & Safe Manipulation},
  author = {Physicl},
  year   = {2026},
  url    = {https://huggingface.co/datasets/physicl/kitchen-workspace-understanding-safe-manipulation}
}