EthanWTL81/geo-obs-ez-data

geo-obs-ez: training split

Rendered videos (SAPIEN) of one object moving by itself in a furnished room, often into another object, for fine-tuning image-to-video models on physically correct contact. Eight tasks, and every motion is meant to go all the way: open a drawer, a door or a lid fully (it starts slightly open), push furniture all the way to the wall into a gap between two props, through a doorway, under a table, turn furniture until it is against the wall, tip an object over. The prompt names what moves and how; it never says how far it gets or whether anything is hit. The scene geometry decides whether the motion is blocked, and the outcomes are balanced.

This repository holds only the training split (base): 39,982 records. Each record is a stereo pair of RGB videos, 49 frames at 24 FPS (about 2 s), 640x480, vertical field of view 0.72 rad. The motion starts at frame 0 and has stopped by frame 42 at the latest.

• render/: the record's camera (left eye).
• render_right/: the right eye, 6.5 cm to the right, parallel axis, same intrinsics.
• meta/: one JSON per record. prompt is the text condition (the training prompt) and instruction a short form of it. outcome is the result. The JSON also holds the cells (planned and realized, and d1), the target and the obstacle (assets, scales, poses), the motion per frame, the contact certificate, the camera (eye, look-at, angles, distance, right eye), the room and the lighting. It holds no absolute paths.

Layout

data/<task>/base_NNN.tar   each holds <task>/render/<name>.mp4, <task>/render_right/<name>.mp4, <task>/meta/<name>.json
manifest/base.csv          one row per record: task, name, outcome, camera / obstacle cell, d1, category, shard, bytes
manifest/shards.json       per shard: records, bytes, sha256
env/flash_attn-2.8.3+cu126torch2.6-cp310-cp310-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl
                           the flash-attn 2.8.3 wheel (torch 2.6, CUDA 12.6, glibc >= 2.28) of the training environment

Unpack every shard into one folder:

for f in data/*/*.tar; do tar -xf "$f" -C /path/to/geo_obs_ez_data; done

A record name reads <prefix>_base_<id>__<camera cell>-<obstacle cell><d1>cm__<target category>__<OUTCOME>, for example odez_base_b00012__side-edge+2.1cm__chest_of_drawers__BLOCK.

• Camera cell. The horizontal angle between the motion and the line of sight: parallel < 30 deg <= side < 60 deg <= perpendicular. The camera looks down moderately (10-45 deg, per task).
• Obstacle cell. d1 is the obstacle's signed distance into the motion's way, along the task's one blocking axis (> 0 = in the way).
  • in: d1 > 3.87 cm, always blocked.
  • edge: |d1| <= 3.87 cm, half blocked, half not.
  • off: d1 < -3.87 cm, never blocked.
• Outcome.
  • BLOCK: the motion stops at the first contact (within 2 mm), short of its end.
  • REACH: the motion completes; it never comes within 2 mm of the obstacle.
  • lay_down has BLOCK (the object ends leaning on the prop) and FLAT (it lies flat on the table).

Every task's records are spread evenly over the 3 x 3 camera x obstacle cells.

task	prefix	records	BLOCK	REACH	FLAT
open_drawer	odez	5,000	2,501	2,499	0
open_hinge	ohez	4,999	2,500	2,499	0
inv_obs	ioez	5,000	2,501	2,499	0
rotate	roez	5,000	2,501	2,499	0
arrange_furniture	afez	4,999	2,500	2,499	0
through_door	tdez	5,000	2,501	2,499	0
under_table	utez	4,989	2,492	2,497	0
lay_down	ldez	4,995	2,498	0	2,497

Terms

Research use only, non-commercial. The rendered objects come from PartNet-Mobility (SAPIEN), Artiverse (which collects models from sources such as Objaverse / Objaverse-XL, 3D-FUTURE and Amazon Berkeley Objects), YCB, ReplicaCAD and procedurally generated books. Several of these sources allow non-commercial research use only, and every object keeps its original license. This repository contains renders only, no 3D models.