PARCEL Stage-1 data: Push-T in Genesis
Data and early training runs of PARCEL (Particle-level Causal Reasoning Learned from Video), code at
https://github.com/jiaka1chen/PARCEL (packed at commit f9c45f0, 2026-11-02). A Panda arm with a stick pushes a T
block on a table, simulated in Genesis 1.4.5; every episode has RGB video, ground-truth 2-D point tracks with
visibility, simulator contacts, poses and actions. DINOv3 features sampled at the tracks are included.
Files
| File | Size (approx.) | Contents |
|---|---|---|
v0_panda.tar |
1.4 GB | 480 episodes: episodes.jsonl, run.json, shard_XXX/ep_XXXXXX.{mp4,npz}, generation logs |
v0_panda_dino.tar.zst |
9.2 GB | DINOv3 ViT-B/16 features at the track positions for every episode, pca.npz |
runs_v1-v10_and_checks.tar.zst |
73 MB | checkpoints (last.pt), args.json, log.jsonl and console logs of the training batches v1-v10 (2026-10-28/29), report media, simulator check reports |
*.tar.list |
small | file listing of each archive |
SHA256SUMS |
small | checksums of the archives |
Extract into the repository's data/ layout:
sha256sum -c SHA256SUMS
mkdir -p data/genesis_pusht
tar -xf v0_panda.tar -C data/genesis_pusht
zstd -dc v0_panda_dino.tar.zst | tar -xf - -C data/genesis_pusht
zstd -dc runs_v1-v10_and_checks.tar.zst | tar -xf - -C data # -> data/runs, data/checks
v0_panda
- 480 episodes, ≤ 20 s each, modes: expert 327, noisy 77, suboptimal 50, wander 26 (expert success ≈ 81 %: coverage of the goal T ≥ 0.9). Validation split used by the code: episode id % 10 == 9.
- Scene: Push-T T block (20 x 20 cm, 4 cm high, 0.35 kg, friction 0.5), Panda + 12 mm x 120 mm stick, fixed camera (elevation 55°, distance 1.2 m, fov 40°), 512 x 512 RGB. Physics 120 Hz, control 10 Hz (action = stick-tip target in the table plane), render 30 fps.
- Particles (K = 180, fixed body-frame points): 120 on the T block (body 0), 24 on the stick and Panda links 6/7 (body 1), 36 on a 6 x 6 table grid (body 2).
ep_XXXXXX.npz:p [F,K,2]pixel tracks,z [F,K]camera depth,vis [F,K]visibility (depth + segmentation test) at 30 fps;local,body,links,actor,seg_entityparticle meta;block_pos/quat,links_pos/quat(30 fps; a free body's pose is Genesis' link frame, offset (0, -0.0321) m from the MJCF origin in the T's body frame);c_flag/c_pos/c_nrm/c_forcestick-block contacts at 120 Hz;action,tip,coverageat 10 Hz;seg(10 Hz);depth(10 Hz, every 20th episode); cameraK,E(OpenCV, world -> camera);goal_pose,block_mass,block_friction,mode,success,first_success,dt,render_every,control_every.ep_XXXXXX.mp4: the RGB video (30 fps).- The archive also holds 65
.mp4files without an.npz(e.g.shard_000/ep_000176-...): partial videos of batches whose generation process was stopped on 2026-10-28. They are not inepisodes.jsonland are never read.
v0_panda_dino
facebook/dinov3-vitb16-pretrain-lvd1689m at 448 x 448 (28 x 28 patch tokens). Per episode patch [F,K,64] (the
patch grid bilinearly sampled at each track), ctx [F,K,64] (the grid 5 x 5 mean-pooled, sampled the same way), both
fp16 and projected on one PCA basis (pca.npz: mean, 64 components, 90.2 % of the variance, fitted on every 10th
episode at 10 Hz), and cls [F,768]. All 30 fps frames. DINOv3 is under Meta's DINOv3 license.
Runs v1-v10
Training batches of the first PARCEL world model on v0_panda (settings, results and analysis per batch:
docs/PLAN.md in the code repository). Checkpoints load with parcel.model.load_model (older ones are the parcel
family). Not included: smoke-test outputs.
Provenance
Generated on cig-04 (RTX A6000, driver 550, torch 2.11 + cu126, genesis-world 1.4.5) with
scripts/gen_pusht.py --mix v0 and scripts/extract_dino.py of the code repository.
- Downloads last month
- 298