D-JEPA decision supervision v1.5
This is an incremental release for RoboTwin robotic manipulation and autonomous driving. It contains no v1 data. Download v1 separately for PushT, Reacher, Granular, unseen-shape PushObj and PushT visual-condition supervision. Physical-robot supervision is planned for v2.0 and is not included here.
Contents
robotwin/candidate_pools/: compact roller-grasping supervision: 27 starts and 99 executed candidates, separated into candidate development, gate development and gate validation. Parameter sweeps, historical expert-conditioned variants and repeated recordings are omitted.robotwin/index.csv: collection, task, seed and shared start-group identities.driving/fit/,driving/calibration/,driving/test/: 200 / 50 / 200 scenes, each with 32 trajectories, predictive features and simulator-scored outcomes.POPULATIONS.json: exact record counts and coverage;SCHEMA.md: field meanings.manifest.json: SHA256 and byte length of each released payload file.SOURCE_HASHES.json: original source identities for transformed small records.
Start here
import numpy as np
root = "D-JEPA-supervision-v1.5"
path = f"{root}/robotwin/candidate_pools/gate_validation/seed-100056"
with np.load(f"{path}/inputs.npz", allow_pickle=False) as inputs:
candidate_ids = inputs["candidate_ids"]
actions = inputs["actions"]
with np.load(f"{path}/labels.npz", allow_pickle=False) as labels:
observed = labels["observed"]
success = labels["success"]
The driving directories retain the cache structure consumed by
djepa.driving.learn.load_cache(path, role) in the
code repository.
All arrays load with allow_pickle=False.
Coverage and reuse
This release provides the available cached supervision and recorded actions; its population counts describe these files, not an expanded reconstruction of every aggregate paper table. Successful and unsuccessful candidates are both retained. Use task and seed groups when constructing new splits. The pre-existing gate development and validation sets remain separate.
Candidate-pool inputs and outcome labels are separate. Camera observations, upstream raw datasets, simulators, map assets and model weights are not included in this lightweight release. Driving query features describe candidate proposals; they are not decoded future images or explicit future-scene latents.
RoboTwin and driving records originate from their respective simulation and upstream model workflows. Their applicable upstream data/resource terms remain in effect; the code repository's license does not independently relicense those resources. See the source identities accompanying each collection.