SpatialCasa-GR1

SpatialLIBERO ported to the Fourier GR1 humanoid on the 12 RoboCasa GR1 tabletop pick-and-place tasks. Every scene has three identical cans and three containers; the instruction names one can by a relation and a goal by another, for example "pick up the can to the left of the pan and place it to the right of the pan". Observations and actions use the GR00T RoboCasa GR1 tabletop format, so a fine-tuned GR00T checkpoint runs in the harness unchanged.

folder	content
eval/	5700 eval episodes (eval_seen 3120, eval_novel 2580) in the SpatialLIBERO levels: scene files for closed-loop rollouts
train/	12 LeRobot v2.0 datasets, one per task, 1000 demos each (200 clean, 100 per sub-factor): 12,000 episodes, 3,178,258 frames, 16 GB

Download

The repository is private; log in with a token that can read it.

pip install -U "huggingface_hub[cli]"
huggingface-cli login

# everything (about 17 GB)
huggingface-cli download twanghcmut/SpatialCasa-GR1 --repo-type dataset --local-dir SpatialCasa-GR1

# eval set only (1.2 GB)
huggingface-cli download twanghcmut/SpatialCasa-GR1 --repo-type dataset --include "eval/*" --local-dir SpatialCasa-GR1

# all 12 train datasets only
huggingface-cli download twanghcmut/SpatialCasa-GR1 --repo-type dataset --include "train/*" --local-dir SpatialCasa-GR1

# one train task (about 1.4 GB)
huggingface-cli download twanghcmut/SpatialCasa-GR1 --repo-type dataset \
  --include "train/gr1_spatial.PlateToPan/*" --local-dir SpatialCasa-GR1

Eval set

An eval episode is a scene file. The harness rebuilds the scene, runs the policy in closed loop and scores the result. The scripted planner solved every episode before release.

Splits

split	cells	cells per task	episodes	base scenes
eval_seen	train cells (reference and goal pairs that appear in training)	5	3120 (260 per task)	new layouts, styles 6 to 11; seen uses training layouts
eval_novel	novel cells (pairs never seen in training)	5	2580 (215 per task)	new layouts, styles 6 to 11

A cell is a (reference, goal) pair such as left_of__right_of. Each task has 20 train cells and 10 novel cells; eval/manifest.json lists them. New eval layouts use seeds from 1000 (eval_seen) and 2000 (eval_novel); train layouts use seeds below 1000.

Robustness factors and levels

There are 8 sub-factors in 4 dimensions:

dimension	sub-factor	what changes
VSR (visual)	background	counter, wall and floor style, and light level
VSR	clutter	distractor objects on the counter
SSU (language)	synonym	paraphrased instruction
SSU	flip	edited instruction whose answer is another can
OCD (distance)	gap_near	target can moved away from its anchor, near bin
OCD	gap_far	target can moved away from its anchor, far bin
SCR (scene)	extra_can	one or two more cans
SCR	anchor_move	the anchor moves so another can becomes the answer

Values per level (SpatialLIBERO Table 2, cans instead of bowls):

change	training	easy	medium	hard
near gap (cm)	0.5-4	5-6.5	8-12	12-16
far gap (cm)	4-8	8-10	12-16	16-20
light (x nominal)	0.8-1.25	0.7-0.8, 1.25-1.4	0.35-0.55, 1.5-1.7	0.27-0.35, 1.7-1.95
texture	pool A	A, new combination	B on half	pool B
clutter objects	1-2, A	3, A	4-5, half B	5-6, B
synonym	pool A	A, unseen sentence	pool B	pool B
flip	relation swap or reference inversion	same kinds	+ goal change or pool-B wording	any edit, pool-B wording
extra cans	0-1	+1	+2	+2
anchor moved	training rule	same rule	one or both anchors	>= 14 cm

One layout under each sub-factor alone at the easy, medium and hard value:

The levels follow the SpatialLIBERO protocol. Every base scene holds one clean episode and one episode per sub-factor, the primary change:

level	per base scene	values
seen	a training layout; each sub-factor alone	training
easy (L1)	each sub-factor alone	easy
medium_single (L2)	each sub-factor alone, on the medium base scene	medium
hard_single (L3)	each sub-factor alone, on the hard base scene	hard
medium	the primary with one companion	medium
hard	the primary with two to four companions (3, 4 and 5 changes equally often)	hard

L1, L2 and L3 apply one change at increasing strength; medium and hard compose it with companions, and their L2 and L3 counterparts are the same episodes with the companions stripped. Companions follow a seeded schedule so every sub-factor is an equally frequent companion of every primary. gap_near and gap_far are never combined, nor is flip with anchor_move, nor a gap with flip or anchor_move.

One base scene per level (green ring: the can to pick; red ring: the answer before the change):

Every episode was solved by the scripted planner. A recipe it could not solve on its base scene was built, with its L2 or L3 counterpart, on another base scene of the same cell and level; a seen item may take another training layout of the task. slot names the base scene group, layout the scene. Check of all groups:

level	rule	groups
all	1 clean + each sub-factor once as primary, every change at the level's value, no excluded pair, every episode solved	420 / 420
seen, easy	primary alone	180 / 180
medium	primary + 1 companion; L2 single for every primary on the same base	120 / 120
hard	primary + 2 to 4 companions; L3 single for every primary on the same base	120 / 120

A re-run of 264 sampled episodes (2 per split, task and level) through the harness with the scripted planner solved 262; one of the two misses passed on a second run.

Episode counts (a combined episode counts toward each of its dimensions):

split	seen	easy	L2	L3	medium	hard	clean	VSR	SSU	OCD	SCR
eval_seen	540	540	480	480	540	540	240	1211	1089	908	1078
eval_novel	-	540	480	480	540	540	180	1082	963	789	972

Tasks

The target container of the RoboCasa task is the main anchor; containers are scaled by 0.6 in every task.

task	anchors
CuttingboardToPan	pan, cutting board, bowl
CuttingboardToPot	pot, cutting board, bowl
CuttingboardToCardboardBox	cardboard box, cutting board, bowl
PlacematToBowl	bowl, placemat, cutting board
PlacematToPlate	plate, placemat, bowl
PlateToBowl	bowl, plate, cutting board
PlateToPan	pan, plate, bowl
PlateToPlate	plate, second plate, bowl
PlateToCardboardBox	cardboard box, plate, bowl
TrayToPlate	plate, tray, bowl
TrayToPot	pot, tray, bowl
TrayToCardboardBox	cardboard box, tray, bowl

Files

eval/
  manifest.json        observation and action spec, success rule, levels, dimensions, counts, cells per task
  episodes.jsonl       one line per episode (5700 lines)
  scenes/<split>/<Task>/<cell>__<layout>__<level>__<recipe_id>.json
  previews/<split>/<Task>/<same name>.png
assets/                images used in this card

A line of episodes.jsonl:

{"episode_id": "eval_seen/CuttingboardToPan/left_of__right_of/s1100/medium/m_flip__background",
 "split": "eval_seen", "slot": "eval_seen/CuttingboardToPan/left_of__right_of/medium",
 "task": "CuttingboardToPan", "cell": "left_of__right_of", "reference": "left_of", "goal": "right_of",
 "layout": "s1100", "level": "medium", "primary": "flip", "recipe_id": "m_flip__background",
 "sub_factor": "flip+background", "sub_factors": ["flip", "background"],
 "op_levels": {"flip": "medium", "background": "medium"}, "dimension": "SSU+VSR", "dimensions": ["SSU", "VSR"],
 "instruction": "pick up the can to the right of the pan and place it in front of the pan",
 "target": "can_2", "old_target": "can_1", "style_id": 4, "light": 0.35695208360734365, "solvable": true,
 "oracle_steps": 262,
 "scene": "scenes/eval_seen/CuttingboardToPan/left_of__right_of__s1100__medium__m_flip__background.json",
 "preview": "previews/eval_seen/CuttingboardToPan/left_of__right_of__s1100__medium__m_flip__background.png"}

primary is the sub-factor of the column (null for clean), target the can to move, old_target the answer on the base scene, op_levels the value of each change, oracle_steps the planner's step count. Previews are for browsing only; the policy never sees them.

Observation and action

Policy input at every step:

key	shape	content
video.ego_view	256 x 256 x 3 uint8	head camera (egoview) rendered at 640 x 400, zero padded to a square and resized to 256
state.right_arm	7	joint positions
state.left_arm	7	joint positions
state.waist	3	joint positions
state.right_hand	6	finger joint positions
state.left_hand	6	finger joint positions
annotation.human.task_description	str	the instruction
annotation.human.coarse_action	str	"unlocked_waist: <instruction>", the GR00T prompt format

The action is the same five parts as absolute joint targets at 20 Hz, as a dict of action.<part> keys or one flat 29-vector (right arm [0:7], left arm [7:14], waist [14:17], right hand [17:23], left hand [23:29]). Action chunks are played one step at a time. The camera is on the head and follows the waist.

Success rule

At most 720 steps (36 s), checked every 10 steps. Success needs all of:

1. the target can satisfies the goal relation,
2. it stands within 15 degrees of upright,
3. no other object moved more than 2 cm,
4. the robot has let go of it and it is at rest (speed below 0.05 m/s).

Report success per split, and per level and dimension within a split; L1, L2 and L3 per sub-factor show the effect of strength alone.

Running the benchmark

The harness is the spatialcasa package in robocasa/ of SpatialLIBERO (branch robocasa-gr1-counter-pnp; robosuite 1.5.1, MuJoCo 3.2.6, setup in robocasa/README.md).

# check the harness with the scripted planner
python -m spatialcasa.bench --eval-dir SpatialCasa-GR1/eval --policy oracle --splits eval_seen --limit 20 --out res/oracle

# a GR00T policy served by Isaac-GR00T's inference service
python -m spatialcasa.bench --eval-dir SpatialCasa-GR1/eval --policy gr00t://localhost:5555 --out res/gr00t

# any policy class: obs dict in, action out; --shard i n splits the run over n processes
python -m spatialcasa.bench --eval-dir SpatialCasa-GR1/eval --policy my_pkg.policy:MyPolicy --shard 0 8 --out res/mine

flag	meaning
--eval-dir	the downloaded eval/ folder
--policy	oracle, gr00t://host:port, or module:Class
--splits, --tasks	restrict the run, default all
--limit	at most this many episodes
--shard i n	run every n-th episode starting at i
--videos k	save mp4 rollouts of the first k episodes
--out	results folder

A custom policy is a class with an optional reset(env, rec) and __call__(obs):

import numpy as np

class MyPolicy:
    def reset(self, env, rec):
        pass

    def __call__(self, obs):
        image = obs["video.ego_view"]
        text = obs["annotation.human.task_description"]
        return np.zeros(29)

Results go to results_<i>of<n>.jsonl (a stopped run resumes) and summary.json (rates by split, level, dimension, sub-factor and task).

Train set

12 LeRobot v2.0 datasets, one per task. Every episode is a scripted demo that passed the eval success rule, on the 20 train cells, background styles 0 to 5, with sub-factors at the training level (milder than eval easy).

Each task follows the SpatialLIBERO training mix: 200 clean demos and 100 per sub-factor, so clean, VSR, SSU, OCD and SCR are 20% each. Demos of each kind are spread round-robin over the 20 cells; sub_factor in meta/episodes.jsonl gives the kind.

task	clean	background	clutter	synonym	flip	gap_near	gap_far	extra_can	anchor_move
each of the 12 tasks	200	100	100	100	100	100	100	100	100
total	2,400	1,200	1,200	1,200	1,200	1,200	1,200	1,200	1,200

task	episodes	frames	mean length	demos per cell (min to max)	instructions
CuttingboardToPan	1000	267,060	267	40 to 57	95
CuttingboardToPot	1000	265,474	265	36 to 57	94
CuttingboardToCardboardBox	1000	261,080	261	26 to 69	89
PlacematToBowl	1000	265,035	265	20 to 60	94
PlacematToPlate	1000	267,381	267	27 to 58	90
PlateToBowl	1000	264,386	264	39 to 57	96
PlateToPan	1000	265,794	266	38 to 56	93
PlateToPlate	1000	267,992	268	43 to 56	92
PlateToCardboardBox	1000	262,375	262	26 to 64	98
TrayToPlate	1000	266,417	266	26 to 59	92
TrayToPot	1000	263,876	264	27 to 60	95
TrayToCardboardBox	1000	261,388	261	23 to 67	93
total	12,000	3,178,258	265

Episodes last 211 to 328 steps at 20 Hz. instructions counts distinct instructions.

Layout of one dataset

train/gr1_spatial.PlateToPan/
  meta/
    info.json            LeRobot v2.0 info: features, fps 20, chunk size 1000, paths
    modality.json        the five state/action slices, the video key and the annotation keys (GR00T reads this)
    episodes.jsonl       per episode: tasks, length, cell, sub_factor, target can, style, light, objects, source demo
    tasks.jsonl          task_index -> text, both "unlocked_waist: <instruction>" and the plain instruction
    stats.json           mean, std, min, max, q01, q99 of observation.state and action (GR00T format)
    relative_stats.json  statistics of the relative arm actions used by GR00T N1.7
  data/chunk-000/episode_000000.parquet ... episode_000999.parquet
  videos/chunk-000/observation.images.ego_view/episode_000000.mp4 ... episode_000999.mp4

Videos are 256 x 256 H.264 at 20 fps, rendered the same way as the eval harness. action holds absolute joint targets; the annotation columns index meta/tasks.jsonl.

Fine-tune GR00T

The datasets load with Isaac-GR00T N1.7 and the ROBOCASA_GR1_TABLETOP embodiment. stats.json and relative_stats.json are already computed for it.

git clone https://github.com/NVIDIA/Isaac-GR00T && cd Isaac-GR00T
# one task
NUM_GPUS=8 MAX_STEPS=60000 GLOBAL_BATCH_SIZE=512 SAVE_STEPS=2000 uv run bash examples/finetune.sh \
  --base-model-path nvidia/GR00T-N1.7-3B \
  --dataset-path /data/SpatialCasa-GR1/train/gr1_spatial.PlateToPan \
  --embodiment-tag ROBOCASA_GR1_TABLETOP \
  --output-dir /tmp/spatialcasa_gr1_plate_to_pan

These are NVIDIA's settings for the GR1 tabletop tasks. Serve the checkpoint with gr00t/eval/run_gr00t_server.py and evaluate with --policy gr00t://host:port.