AdaDexTrack resources

Resources for AdaDexTrack: Dynamic Modulation for Adaptive and Generalizable Dexterous Manipulation Tracking, CVPR 2026.

Code and instructions · Project website · Paper

Files

File	Restored paths	Contents
assets.tar.gz	assets/	Object meshes and URDFs, xArm/LEAP robot model, floor texture, object latents
references.tar.gz	data/adadextrack_full/	All 2,765 reference motions, manifests, semantic groups, split lists and provenance
bc_tracker_unseen_ref.pth	checkpoints/bc_tracker_unseen_ref.pth	Pretrained unseen-trajectory BC tracker, 273 observations and 22 actions
bc_tracker_unseen_obj.pth	checkpoints/bc_tracker_unseen_obj.pth	Pretrained unseen-object BC tracker, 273 observations and 22 actions
modulator_unseen_ref.pth	checkpoints/modulator_unseen_ref.pth	Pretrained unseen-reference modulator, paired with bc_tracker_unseen_ref.pth
modulator_unseen_obj.pth	checkpoints/modulator_unseen_obj.pth	Pretrained unseen-object modulator, paired with bc_tracker_unseen_obj.pth
object_ae.pth	checkpoints/object_ae.pth	Optional original object AE, with encoder, decoder and training state
object_ae_data.tar.gz	data/object_ae/	Optional 57 input meshes, 55/2 train/validation split and SHA256 manifest
example_specialists.tar.gz	checkpoints/specialists/examples/	Optional pretrained stamp, toothpaste and wineglass specialists (groups 322, 369, 435), configs and metrics
release.json	Download metadata	Archive and per-file SHA256 checksums, sizes and coverage

The corpus contains 2,765 motions, 50 objects, 444 semantic groups, with 300 frames per motion. Reference IDs are 0–2764; group IDs are 0–443. Each motion is stored once and shared by both experimental partitions.

Partition	Train	Test
Unseen trajectory	2,212	553
Unseen object	2,520	245

This package contains the full corpus, without height, checkpoint-success or accuracy filtering. References were generated with DiffH2O and retargeted to the robot. Their dictionaries contain object poses, robot joint states and link transforms. Position units are meters, angles are radians and quaternions are xyzw. The original NumPy dictionaries use pickle serialization; the project loader reads these project resource files with allow_pickle=True.

Two BC trackers are released: bc_tracker_unseen_ref.pth is trained on the 2,212-reference unseen-trajectory training split and is evaluated on unseen_trajectory_test.csv (553 references); bc_tracker_unseen_obj.pth is trained on the 2,520-reference unseen-object training split and is evaluated on unseen_object_test.csv (245 references). Both have 273 observations and 22 actions. Each is intended as the frozen tracker for its corresponding modulator experiment. Semantic specialists for the three example groups are available as the optional example_specialists.tar.gz component; full-split specialists are trained with the code repository's specialist pipeline. The two pretrained modulators are included as modulator_unseen_ref.pth and modulator_unseen_obj.pth, each with 273 observations and 202 actions. Use each with its matching released BC tracker and the fixed object latents.

Download

From the code repository root, in the configured Python environment:

pip install 'huggingface_hub>=0.13.4'
python scripts/download_resources.py

The downloader resolves one Hub commit, verifies the archive and individual-file checksums, and restores the paths above. It preserves identical installed files and fails on conflicting local files. For a reproducible version, pass --revision <full Hub commit hash>.

To download only the meshes, robot models, textures and object latents:

python scripts/download_resources.py --components assets

For manual installation, extract both .tar.gz archives into the code repository root and put both bc_tracker_unseen_ref.pth and bc_tracker_unseen_obj.pth, plus modulator_unseen_ref.pth and modulator_unseen_obj.pth, in checkpoints/. Archives already include their top-level assets/ and data/adadextrack_full/ paths. Isaac Gym Preview 4 must be obtained and installed separately.

Optional example specialists

python scripts/download_resources.py --components example_specialists
python scripts/collect_specialists.py --specialist-root checkpoints/specialists
python scripts/train_bc.py

The three example specialists use groups 322 (stamp, two references), 369 (toothpaste, four references), and 435 (wineglass, one reference). Each was trained for 2,001 PPO epochs. Their mean-error success rates are all 100%; framewise Completion rates are 100%, 79.67% and 100%, respectively, using seed 42 and full 300-frame evaluation without initial XY/yaw perturbations. Per-reference metrics and training settings accompany each checkpoint. This component is opt-in and does not overwrite locally trained specialists. See the code repository's docs/specialists.md for inference commands.

Optional object autoencoder

The default download installs assets, references, both BC trackers and both modulators. AE weights and AE input meshes are opt-in:

python scripts/download_resources.py --components object_ae object_ae_data

The original AE best checkpoint contains a 64-dimensional PointNet encoder and 2,048-point decoder, plus Adam/scheduler state. Its recorded budget is 1,000 epochs; it was selected after epoch 925. The 57 input meshes are split into 55 training inputs and two validation inputs: rubberduck and flute. data/object_ae/manifest.json records the split and per-file hashes. These inputs retain the source mesh terms described below.

The fixed 50 policy latents in assets/latents remain unchanged. Newly seeded sampling/encoding does not promise to reproduce their bytes. See the code repository's docs/object_ae.md for AE train, encode, decode and representation compatibility. New representations require consistently retrained policies.

Resource terms

There is no single blanket license over this mixed resource bundle. The code repository's BSD license does not replace third-party asset terms. See RESOURCE_LICENSES.md for source information and restrictions, and notices/ for the retained available license files.

In particular, GRAB's published terms include noncommercial-use and redistribution restrictions. Including files here does not grant additional rights or establish permission from the original rightsholders. Generated references and trained weights are distinct resources; no blanket relicensing of third-party material is asserted.

Evaluation

Run each tracker with its matching test split from the code repository root:

python scripts/infer_bc.py \
  --checkpoint checkpoints/bc_tracker_unseen_ref.pth \
  --ref-csv data/adadextrack_full/splits/unseen_trajectory_test.csv \
  --batch-size 553 --device cuda:0 --seed 42

python scripts/infer_bc.py \
  --checkpoint checkpoints/bc_tracker_unseen_obj.pth \
  --ref-csv data/adadextrack_full/splits/unseen_object_test.csv \
  --batch-size 245 --device cuda:0 --seed 42

Add --render for a viewer. python scripts/download_resources.py downloads both BC trackers and both modulators by default; use --components bc_tracker_unseen_ref bc_tracker_unseen_obj to download only the two weights.

Run the pretrained modulator with its matching frozen BC tracker:

python scripts/infer_modulator.py --dataset unseen_ref \
  --checkpoint checkpoints/modulator_unseen_ref.pth \
  --tracker-checkpoint checkpoints/bc_tracker_unseen_ref.pth --batch-size 553
python scripts/infer_modulator.py --dataset unseen_obj \
  --checkpoint checkpoints/modulator_unseen_obj.pth \
  --tracker-checkpoint checkpoints/bc_tracker_unseen_obj.pth --batch-size 245

Add --render for a viewer or --split train for the corresponding training set.

The code uses L2 position errors, quaternion angular errors and 16-finger-joint MAE over all 300 frames. Initial object XY/yaw perturbations are disabled. See the code repository for the exact Mean-error and Completion definitions.

@InProceedings{Adalibieke_2026_CVPR,
  author = {Adalibieke, Jianibieke and Han, Qianwei and Liu, Xueyi and Qin, Yuzhe and Yi, Li},
  title = {AdaDexTrack: Dynamic Modulation for Adaptive and Generalizable Dexterous Manipulation Tracking},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2026}
}