Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

Aldarondo 2024 locomotion retargeted to Fetch

This dataset contains the strict locomotion subset of all 38 sessions from the coltrane rat in Aldarondo et al. (2024), deterministically retargeted onto the unmodified 10-DoF Brax v1 Fetch body. It is the data-only input to Demo F in the Embodied SSL + RL workshop.

Each row is a non-overlapping 64-frame trajectory sampled at 50 Hz. Rows contain only the derived Fetch trajectory and source provenance; raw rat keypoints, neural recordings, behavior labels, and video are not redistributed.

Data fields

Each data/<split>/<session>.npz shard contains:

field shape dtype meaning
joint_angles (N,64,10) float32 Fetch joint angles, radians
root_position (N,64,3) float32 Fetch root trajectory, Fetch model units
root_quaternion (N,64,4) float32 scalar-first root quaternion
feet_local (N,64,4,3) float32 realized Fetch foot endpoints in root frame
contacts (N,64,4) uint8 source-derived stance mask
command (N,3) float32 hindsight forward/lateral displacement and yaw
source_start (N,) int32 first frame in the named source session
source_speed_mps (N,) float32 net rat displacement speed; may be small on curved paths
source_path_speed_mps (N,) float32 mean rat path speed
ik_foot_rmse (N,) float32 semantic endpoint fitting error
contact_speed_mean (N,) float32 realized Fetch stance-foot speed
minimum_foot_height (N,) float32 minimum realized foot height
joint_limit_fraction (N,) float32 fraction within 1% of a joint limit

Foot ordering is front-right, front-left, back-right, back-left. Joint ordering and every transformation parameter are recorded in manifest.json.

Construction

The builder applies Demo B's frozen geometric locomotion screen to non-overlapping 64-frame blocks: mean path speed above 0.10 m/s, coordinated joint motion, less than 90 degrees turning, and limited neck-height drift. Retargeting then uses a smoothed trunk frame and four semantic paws, body-size normalization, stance detection/pinning, and bounded sequence-level inverse kinematics with pose/velocity/acceleration regularization. Rodent bone rotations are not copied.

Train, validation, and test splits contain 26, 6, and 6 entire recording sessions. No session appears in more than one split.

Intended use and limitations

The dataset supports conditional motion modelling, motion-prior diagnostics, and simple control experiments on Fetch. It represents rodent-derived Fetch motion, not rat biomechanics: morphology, length, mass, actuation, and physical feasibility differ. The trajectories are kinematic IK results and are not torque supervision or guaranteed dynamically feasible demonstrations.

Source, attribution, and license

The source recordings are:

Aldarondo, D. et al. A virtual rodent predicts the structure of neural activity across behaviours. Nature 632, 594–602 (2024). https://doi.org/10.1038/s41586-024-07633-4

Original dataset: https://doi.org/10.7910/DVN/FB0MZT

The Harvard Dataverse source declares the Open Data Commons Attribution License (ODC-By) 1.0. This derived dataset is redistributed under ODC-By 1.0 and retains the required source attribution. Users should also cite the original paper and dataset. The Fetch implementation originates in Brax, licensed under Apache 2.0.

Reproduction

The exact schema, source hashes, code hashes, configuration, per-session counts, quality gates, and shard SHA-256 values are in manifest.json. Builder and validator source live in the companion workshop repository under demo_f/dataset/.

Release statistics

Schema 1.0.0 contains 2,903 clips: 2,156 train, 335 validation, and 412 test. Generated statistics are authoritative in manifest.json.

Downloads last month
38