DeformPath2 and AltDynamics

This repository contains two related datasets for deformable-object manipulation with synchronized RGB-D point clouds and motion-capture tool poses:

• DeformPath2: the main multi-episode dataset of dough-shaping demonstrations.
• AltDynamics: an alternative dynamics recording exported from the snimanje_16_6 source bag, including interpolated point clouds, tool trajectories, temporal annotations, and STL trajectory visualizations.

The recordings track two tools:

• UR5e_spathla
• gen3_spathla

The tool poses are recorded in a motion-capture coordinate frame. Point clouds contain the dough and surrounding scene points after the dataset-specific filtering steps.

Dataset overview

Dataset	Main content	Interpolated samples	Temporal annotations
DeformPath2	Multiple dough-shaping episodes	45 complete interpolated episodes in the release subset	Available for selected source episodes
AltDynamics	Alternative dynamics recording	534 aligned samples	Episode 18: 28 motions; Episode 3: 41 motions

The original DeformPath2 release also contains RGB videos, depth archives, metadata, meshes, and dataset statistics. AltDynamics contains the raw export, interpolated export, source ROS 2 bag directories, and visualizations for the available annotated recordings.

Repository layout

DeformPath2_release/
├── README.md
├── DeformPath2_interpolated/
│   └── snimanje_23_10/
│       └── episode*_kugla/
│           ├── paths_interpolated.pt
│           ├── pointclouds_interpolated.pt
│           └── sequence_metadata.json
├── AltDynamics/
│   ├── AlternativeDynamicsBag/
│   │   └── snimanje_16_6/
│   │       ├── episode3_dynamics/
│   │       ├── episode11_dynamics/
│   │       └── episode18_dynamics/
│   ├── AlternativeDynamicsExported/
│   ├── AlternativeDynamics_interpolated/
│   ├── visualizations/
│   └── README.md
├── depth/
├── videos/
├── metadata/
├── meshes/
└── visualizations/

DeformPath2

DeformPath2 contains dough-shaping demonstrations recorded with an RGB-D camera and an OptiTrack motion-capture system. The main release includes videos, depth data, filtered point clouds, interpolated tool poses, collision meshes, and dataset-level statistics.

DeformPath2 data formats

Videos

The RGB videos are H.264 MP4 files, normally recorded at approximately 30 Hz.

Depth

Depth archives are compressed NumPy files:

Key	Shape	Type	Description
depths	[N, 480, 848]	uint16	Depth in millimetres; zero marks invalid depth
timestamps_ns	[N]	int64	Frame timestamps
encoding	scalar	string	Usually 16UC1

Interpolated paths

Each paths_interpolated.pt file contains a one-item list with a record:

import torch

record = torch.load(
    "DeformPath2_interpolated/snimanje_23_10/episode20_kugla/paths_interpolated.pt",
    weights_only=False,
)[0]

path = record["path"]
valid = record["stream_validity"]
pose_frames = record["pose_frames"]

The fields are:

Field	Shape	Description
path	[T, 2, 14]	Tool poses and velocities
stream_validity	[T, 2]	Validity mask for each tool
pose_frames	list	Tool names in path-axis order

The 14 path features are:

x, y, z, qx, qy, qz, qw,
vx, vy, vz, wx, wy, wz,
relative_timestamp_s

Interpolated point clouds

Each pointclouds_interpolated.pt file contains a one-item list of cloud frames. Each frame has shape [N, 7]:

x, y, z,
normalized_frame_time,
frame_number,
valid_flag,
relative_timestamp_s

RGB is used by the filtering process where available, but RGB columns are not retained in the interpolated output.

AltDynamics

AltDynamics is an alternative recording exported from the snimanje_16_6 source bag. The interpolated directory contains 534 aligned cloud and pose samples:

AltDynamics/AlternativeDynamics_interpolated/
├── paths_interpolated.pt
├── pointclouds_interpolated.pt
├── sequence_metadata.json
├── camera_info.json
└── floor_estimate.json

Its serialized path and point-cloud formats match the DeformPath2 interpolated format described above.

The source bag contains RGB messages on:

/camera/camera/color/image_raw

and point clouds on:

/camera/camera/depth/color/points

Temporal annotations

Temporal annotations are available for the following AltDynamics recordings:

• episode18_dynamics: 28 annotated motions
• episode3_dynamics: 41 annotated motions
• episode11_dynamics: annotation file present; currently contains one segment

The annotation boundaries use point-cloud ordinals and map exactly to the synchronized RGB frame ordinals for the available recordings. Each segment contains a start ordinal, an exclusive end ordinal, and the corresponding timestamps.

Visualizations

Main DeformPath2 release visualizations

The main release includes:

The STL animation shows the UR5e mesh in blue and the Gen3 mesh in orange, with accumulated tool trajectories.

AltDynamics visualizations

AltDynamics includes trajectory projections, point-cloud statistics, a point-cloud time grid, annotated RGB motion grids, and an STL trajectory animation:

The annotated RGB grids show one motion segment per cell and advance each cell through its own segment during playback.

STL trajectory rendering

The reusable renderer is:

DeformPath2_release/DeformPath/render_stl_trajectories.py

The original DeformPath2 release grid can be rendered with:

cd DeformPath2_release
python DeformPath/render_stl_trajectories.py

To render AltDynamics with the Episode-18 manual candidate alignment:

cd /path/to/deformpath_training
python DeformPath2_release/DeformPath/render_stl_trajectories.py \
  --trajectory-root DeformPath2_release/AltDynamics/AlternativeDynamics_interpolated \
  --manual-alignment \
    /path/to/TaichiDough/experiments/differentiable_mpm/data/episode18_dynamics_manual/episode18_tool_pose_frame0300_alignment.json \
  --camera-orbit-deg 90 \
  --camera-roll-deg 90 \
  --camera-zoom 2 \
  --frames 20 \
  --fps 5 \
  --output DeformPath2_release/AltDynamics/visualizations/episode18_dynamics_stl_trajectory.gif

The supplied alignment is marked manual_candidate and validated_geometry: false. It is applied for visualization and should not be treated as a validated simulator calibration. The alignment was created from a different serialized interpolation export, although both exports use the mocap coordinate frame.

RGB motion-grid generation

The RGB motion grids are generated directly from the ROS 2 bag and temporal_annotations.json. The annotation files currently available in AltDynamics are:

AltDynamics/AlternativeDynamicsBag/snimanje_16_6/episode18_dynamics/temporal_annotations.json
AltDynamics/AlternativeDynamicsBag/snimanje_16_6/episode3_dynamics/temporal_annotations.json
AltDynamics/AlternativeDynamicsBag/snimanje_16_6/episode11_dynamics/temporal_annotations.json

The grid renderer uses the RGB topic and maps each annotated point-cloud segment to the synchronized RGB frame range.

Limitations

• The AltDynamics STL registration is a manual candidate and has not been validated against the exact serialized AltDynamics interpolation file.
• Interpolated point clouds do not retain RGB columns.
• Temporal annotation coverage is incomplete across the available recordings.
• Some release metadata describes the larger source dataset; the files in this repository are the included release subset.

Citation

@misc{deformpath2_2026,
  title={DeformPath2: A Multimodal Dataset for Deformable Object Manipulation},
  author={Antonio Skara},
  year={2026},
  url={https://huggingface.co/datasets/AntonioSkara132/DeformPath2}
}