AnyGoal datasets

paper: https://arxiv.org/abs/2610.26844

AvoidT	StackT	PushT Antigoal	CubeHop
CubeDiagonal	Avoid Peg	ExtractPeg	TwoRoom Route

Datasets for AnyGoal, a compositional generalization suite for world-model research. The files use the Stable World-Model trajectory HDF5 format and can be used directly with AnyGoal's training, rollout, and planning APIs.

This repository contains the eight pre-collected AnyGoal task datasets (400 episodes each) and the Insert Peg base dataset. The upstream PushT, Cube, and TwoRoom base datasets are downloaded separately from their original publishers.

Quick start

Install AnyGoal, then download every dataset published in this repository into its configured data directory:

pip install anygoal

hf download HazelNam/anygoal --repo-type dataset \
  --local-dir "$(anygoal config path data)"

To download only one dataset with the Hugging Face CLI:

hf download HazelNam/anygoal extract_peg_test.h5 \
  --repo-type dataset \
  --local-dir "$(anygoal config path data)"

AnyGoal can also download all published datasets or the dataset/checkpoint pair needed for an evaluation:

anygoal artifacts download --datasets --all
anygoal artifacts download --model driftworld --dataset extractpeg

Artifacts can be downloaded automatically immediately before evaluation:

anygoal eval planning \
  --model driftworld \
  --dataset extractpeg \
  --download

Existing local files are reused. By default, datasets are stored in Stable World-Model's dataset directory (normally ~/.stable_worldmodel/datasets). Set ANYGOAL_DATA_ROOT to use another location.

Dataset catalog

Dataset ID	Family	Canonical filename	Available here
pusht	PushT	pusht_expert_train.h5	No — upstream base data
avoidt	PushT	avoidt_test.h5	Yes
stackt	PushT	stackt_test.h5	Yes
pusht-antigoal	PushT	pusht_antigoal_test.h5	Yes
cube	Cube	cube_single_expert.h5	No — upstream base data
cubehop	Cube	cube_hop_test.h5	Yes
cubediagonal	Cube	cube_diagonal_test.h5	Yes
insertpeg	Insert Peg	manipulator_insert_peg_weak_expert_train.h5	Yes
avoid_peg	Insert Peg	avoid_peg_test.h5	Yes
extractpeg	Insert Peg	extract_peg_test.h5	Yes
tworoom	TwoRoom	tworoom.h5	No — upstream base data
tworoom_route	TwoRoom	tworoom_route_test.h5	Yes

Aliases such as cube_diagonal, extract_peg, and tworoom-route are accepted by the AnyGoal API. Canonical IDs and filenames are recommended for scripts and stored artifacts.

Upstream base datasets

Download the PushT, Cube, and TwoRoom base datasets from their official LeWM repositories:

hf download quentinll/lewm-pusht pusht_expert_train.h5.zst \
  --repo-type dataset --local-dir "$(anygoal config path data)"
unzstd "$(anygoal config path data)/pusht_expert_train.h5.zst"

hf download quentinll/lewm-cube cube_single_expert.tar.zst \
  --repo-type dataset --local-dir "$(anygoal config path data)"
tar --zstd -xvf "$(anygoal config path data)/cube_single_expert.tar.zst" \
  -C "$(anygoal config path data)"

hf download quentinll/lewm-tworooms tworoom.tar.zst \
  --repo-type dataset --local-dir "$(anygoal config path data)"
tar --zstd -xvf "$(anygoal config path data)/tworoom.tar.zst" \
  -C "$(anygoal config path data)"

Use a dataset

Download and open a dataset from Python:

from anygoal.artifacts import download_dataset
from stable_worldmodel.data import HDF5Dataset

dataset_path = download_dataset("extractpeg")
dataset = HDF5Dataset(path=dataset_path, keys_to_cache=["action"])

print(dataset_path)
print(dataset.lengths)

Validate a local file without modifying it:

anygoal data validate \
  "$(anygoal config path data)/extract_peg_test.h5" \
  --dataset extractpeg

Use the canonical dataset ID with AnyGoal evaluation commands:

anygoal eval rollout --model driftworld --dataset extractpeg --seed 0 --dry-run
anygoal eval planning --model driftworld --dataset extractpeg --seed 0 --dry-run

Remove --dry-run after checking the resolved paths and command.

AnyGoal tasks

The eight AnyGoal datasets recombine familiar states and actions into new goals while staying within a base task family:

• AvoidT (avoidt): move the PushT agent to a sampled corner target while treating the physical T as an obstacle.
• StackT (stackt): rotate and move the physical T so its head overlays a visible goal T.
• PushT Antigoal (pusht-antigoal): move the physical T toward an invisible corner target while the usual visible PushT goal acts as a decoy.
• CubeHop (cubehop): move and rotate the end effector between floor targets and elevated waypoints around a cube.
• CubeDiagonal (cubediagonal): solve the single-cube task using a diagonal grasp across opposite vertical cube edges.
• Avoid Peg (avoid_peg): move the manipulator around the target box without contacting the peg.
• ExtractPeg (extractpeg): grasp an inserted peg, extract it, carry it through a waypoint, and release it at a sampled destination.
• TwoRoom Route (tworoom_route): follow one of eight start–waypoint–goal routes without rendering the target.

See the AnyGoal documentation for task details, model integrations, training, and evaluation.

Data format

Each HDF5 file follows the Stable World-Model trajectory contract. action and pixels are required, optional state and task-specific arrays share the same leading time dimension, and episode boundaries are stored with ep_len or episode. AnyGoal datasets also carry an anygoal_dataset_id root attribute for task identity.