AdaRoute Dataset v1

AdaRoute v1 is a matched intervention dataset for learning action-conditioned robot dynamics around novel obstacles. It contains raw LIBERO-10 rollouts from a frozen pi0.5 policy, direct obstacle failures, and motion-planner bridges followed by policy continuation.

Project code and collection details: Loule0-0/AdaRoute.

Release Summary

• 375 matched roots: 300 base roots and 75 controlled retries.
• 1,529 schema-valid HDF5 trajectories.
• 104,247,555,293 bytes on disk (about 98 GiB as reported by du -sh).
• LIBERO-10 tasks 0, 8, and 9 with 526, 494, and 509 trajectories.
• 764 collision-free executed planner bridges.
• 307 successful policy continuations after planning.
• 328 candidate continuations labeled stuck.
• Fixed cylinder, bar, and wall obstacles with varied pose and color.

Matched Branches

Each root shares the same task state, object offsets, and obstacle across:

1. clean_policy: pi0.5 without an obstacle.
2. obstacle_policy: pi0.5 executes directly around the obstacle.
3. candidate_01_shortest: shortest feasible planner bridge, then pi0.5.
4. candidate_02_clearance: highest-clearance bridge, then pi0.5.
5. candidate_03_recovery_prior: stage-compatible endpoint, then pi0.5.

Partial roots are intentionally retained as negative planning-coverage data. They contain fewer than three feasible planner branches, but every stored trajectory is complete.

Stored Signals

Every frame stores:

• observations/agent_rgb and observations/wrist_rgb: dual 256 x 256 RGB.
• observations/*_obstacle_mask: simulator-GT binary obstacle masks.
• proprio/eef_*, gripper_qpos, arm_qpos, and arm_qvel.
• actions/commanded_osc: normalized 7D LIBERO EEF delta pose plus gripper.
• actions/realized_eef_delta and actions/source.
• labels/reward, done, stuck, and obstacle-contact labels.
• privileged/sim_qpos, sim_qvel, object poses, and obstacle pose.

RGB, proprioception, task text, and actions are deployable AC-WM inputs. Simulator state, contacts, reward, and GT masks are offline supervision and evaluation labels only. The obstacle masks are not online trigger inputs.

commanded_osc[i] is the action that produced observation i, rather than the next action from that observation. The release uses schema adaroute_raw_v1; exact pi0.5 query-chunk references are available only in newer raw-v2 collection.

Repository Layout

README.md
manifests/
  adaroute_train_v1.jsonl
  adaroute_train_v1_retries_v4.jsonl
  adaroute_train_v1_with_retries.jsonl
train/
  task_00/<root>/<branch>/{trajectory.h5,metadata.json}
  task_08/<root>/<branch>/{trajectory.h5,metadata.json}
  task_09/<root>/<branch>/{trajectory.h5,metadata.json}
scene_files/
audit_final.json
audit_final.csv
file_index.tsv
sample_episode_inspection.json

Download

hf download Jaqen0-0/AdaRoute \
  --repo-type dataset \
  --local-dir data/adaroute_train_v1

The corpus is intentionally distributed as raw HDF5 rather than flattened into a Hugging Face datasets table. This avoids lossy or duplicated re-encoding of frame-aligned images, masks, physical interactions, and privileged evaluation labels.

Minimal Read Example

import h5py

path = "train/task_00/<root>/<branch>/trajectory.h5"
with h5py.File(path, "r") as episode:
    agent_rgb = episode["observations/agent_rgb"][0]
    wrist_rgb = episode["observations/wrist_rgb"][0]
    action = episode["actions/commanded_osc"][0]
    obstacle_mask = episode["observations/agent_obstacle_mask"][0]

Intended Use

The release supports AC-WM training, action-following evaluation, physical interaction modeling, collision/stuck prediction, planner-endpoint scoring, and policy-planner handoff research. Do not use privileged simulator fields as deployable inputs when reporting real-robot-compatible results.

No standalone license is declared for this dataset release. Users are responsible for complying with the licenses and terms of AdaRoute, LIBERO, MuJoCo, OpenPI, and the underlying simulation assets.