Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 92, in _split_generators
                  inferred_arrow_schema = pa.concat_tables(pa_tables, promote_options="default").schema
                                          ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "pyarrow/table.pxi", line 6268, in pyarrow.lib.concat_tables
                File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
                  return check_status(status)
                File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
                  raise convert_status(status)
              pyarrow.lib.ArrowTypeError: Unable to merge: Field json has incompatible types: struct<schemaVersion: int64, mediaType: string, manifests: list<item: struct<mediaType: string, digest: string, size: int64, annotations: struct<io.containerd.image.name: string, org.opencontainers.image.ref.name: string>>>> vs list<item: struct<Config: string, RepoTags: list<item: string>, Layers: list<item: string>>>
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 68, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

V2D Challenge 2026: Track 2 submission container (team Neural Motion)

This is the Docker image tarball (docker save | zstd) for the award container check described in the submission kit's CONTAINER_SUBMISSION.md. It covers Track 2, Tiers 1, 2 and 3: labels tier1, tier2 and tier3, one per tier, 30 policies each. Each label is the submission counted on all four boards of its tier (Kaggle Final submissions):

Label Track Counted file Kaggle ids (AUC, SP-SR, MP-SR, MPPE)
tier1 track_2_tier1 t1_M_20261104T194704Z.csv 57996985, 57996986, 57996989, 57996990
tier2 track_2_tier2 t2_M_20261104T194709Z.csv 57997002, 57997005, 57997007, 57997013
tier3 track_2_tier3 t3_C_20261104T201650Z.csv 57997757, 57997760, 57997761, 57997764
  • Load: zstd -dc neural-motion-v2d-submission.tar.zst | docker load
  • Verify: docker run --rm --network none -v /path/to/dataset:/data:ro neural-motion-v2d-submission:final /submission/run.sh verify --dataset-root /data
  • Reproduce one tier: docker run --rm --gpus all --network none -v /path/to/dataset:/data:ro -v "$PWD/out:/output" neural-motion-v2d-submission:final /submission/run.sh reproduce --label tier1 --dataset-root /data --output /output --parallel 8 (likewise --label tier2 and --label tier3)
  • Code: https://github.com/NMotion-Initiative/v2d-challenge-2026 at commit 96e6d22ff0e40cc91c9df96ccc5d193aa42b83e1, the commit recorded in every counted file and in /submission/submission.json (Apache-2.0).
  • Runtime: FlashCHORD (nvidia-isaac/video_to_data @ 8c778f9). It needs one NVIDIA GPU with at least 48 GB and driver 580 or newer.
  • Technical report: neural_motion_track2.pdf. It was revised after the freeze, under the same name. Results, scores, ids, links and hashes are unchanged throughout.
    1. Errata (revision 071c3faa): Table 3's margin for Tier 1 episode 9 is 0.086, not 0.087 (the decision is unchanged), and E's saved configuration also differs from C's in total steps and the 9 curriculum thresholds, because E trains with 0.7x the steps.
    2. Appendix update (revision ea5bba2f): Appendix A describes this container as built and checked (verify, full reproduce and reproduced scores), the selector hash, the unused arm E replica runs, and the motivation of the arm E margin.
    3. Correction (revision 2ddf52d8): revision ea5bba2f counted three of E's losses as beyond the margin against M's choice; two of them (episode 17 in both tiers) come from C, which E extends. The main text went back to the submitted wording, and the appendix compares E with M and with C.
    4. Raw-source audit (this version): every claim was traced to raw tool outputs, logs and code (187 claims; no wrong number; 10 imprecise wordings confirmed after an adversarial re-check). Four main-text phrases changed: E's checkpoints save C's objective; the Tier 3 frame check read the frame-0 object pose and human position of two episodes; when M was defined, B's and Tier 2 C's Kaggle scores were known (Tier 1 C was never uploaded); and Tier 3's result points to the reconstruction (an inference). The appendix states the remaining points more precisely. The version submitted at the freeze is in revision 78982f01 of this repository (sha256 b54683602344780432f886dd9e64df3ca65cb9af8e6102208839ae4dfe621c52).

Files

File What it is
neural-motion-v2d-submission.tar.zst The image, neural-motion-v2d-submission:final (docker save, zstd).
neural-motion-v2d-submission.tar.zst.sha256 Its sha256.
container_summary.txt Image id, base image, commit, tarball sha256 and size.
hostside/py_shim.sh, hostside/t3_manifest_patch.py The two host-side staging tools described below.

Provenance of this image

  • /submission/code is a clean git clone of github.com/NMotion-Initiative/v2d-challenge-2026, checked out detached at commit 96e6d22ff0e40cc91c9df96ccc5d193aa42b83e1, with no modified or untracked files (/submission/run.sh verify checks this). /submission/run.sh and the Dockerfile are that commit's container/ files, unchanged. /submission/README.md is that commit's container/README.md with its title extended to Tier 3 and a Tier 3 section appended.
  • The image was staged with that commit's container/assemble.sh and built with its container/build_and_verify.sh, on the VM that packed Tiers 1 and 2, from its flash_chord:8c778f9 image. Three host-side steps ran while staging. They are not in the commit, and nothing of them is in /submission/code:
    1. hostside/py_shim.sh (sha256 482f9b415e7684235ed967048d6e78681cf68f6bf0ff4fac14da081ea6a8e994): the commit's container/make_submission_json.py maps only tiers t1 and t2 to labels, so it stopped on Tier 3. The wrapper runs that file unchanged with the one mapping t3 -> tier3 added. Every check it makes is unchanged, and every other command passes through.
    2. hostside/t3_manifest_patch.py (sha256 2be1f7254a261f2254a748b32aa67d4a677bec7e789297a4f967b83d65f0a5c1): the commit's manifest writer records no frame_references. After staging, this step copied the Tier 3 frame references byte for byte to /submission/frame_references/tier3. They are our monocular reconstruction, the --frame-references input that packed the Tier 3 CSV: 30 episode_XXXXXX.npz files, sha256sum episode_*.npz | sha256sum = 7402ce549d2623f19108fc674d0f75566a2a56d181d74138dc770cf03fcad42c. The step then added "frame_references": "frame_references/tier3" to the tier3 entry of /submission/submission.json.
    3. The Tier 3 section of /submission/README.md. Nothing else in the manifest or the image changed.
  • Checkpoint selection: the Tier 1 and Tier 2 selection logs record the selector hash prefix 91b27eee, i.e. v2d/select_checkpoints.py at the launch commits (0d605e4, 3c85936, fe1da6c, c9440d2). At 96e6d22 the file hashes to 548bdc8d, because commit 8ff5e89 added a branch taken only for tier t3; the Tier 1/2 code path is unchanged (git diff 3c85936 96e6d22 -- v2d/select_checkpoints.py). Selection is not re-run in this image; it only explains how each checkpoint was chosen.
  • Motions: each tier's retargeted motions and support stages sit at the absolute paths the checkpoints recorded (/opt/v2d/<tier>/assets). The policies were trained on several VMs; before staging, every training VM's copy of those trees (from its disk snapshot) was compared file by file with the staged copy: all identical.
  • Checked before release, with --network none:
    • before the build, the staged verifier's own re-pack of every label matched the submitted file (tier3 with the staged frame references);

    • run.sh verify --dataset-root /data: 29 pass, 0 fail, 0 warn, 0 skip (every hash, the commit, the 30 checkpoints of each label, and the re-pack of all three labels);

    • a 2-episode smoke reproduce of each label: provenance identical to the stored rollouts for all 6 episodes;

    • a full reproduce of every label on the image loaded from this tarball (docker load reports image id sha256:c80b68ee1bcbe400c5e585171b553ecd7fdfefb56866454d238acee677c634ad): 30 episodes each, world_count=1, --parallel 8, about 4.5 min per label on 8 H100s. Then run.sh verify --reproduced for all three: 32 pass, 0 fail, 0 warn, 3 skip (the 3 skips are the re-packs, which need --dataset-root and passed in the verify above). Each reproduced file has the submitted rows and provenance identical to the submitted file. Reproduced object positions vs submitted: median 0.87, 1.28 and 1.28 cm (tier1, tier2, tier3); the long tail (95th percentile 99-126 cm) comes from episodes whose object is dropped in one draw and not in the other, as the simulation is not bitwise deterministic.

    • Scores of the reproduced files, as the packer prints them (clean reference), next to the leaderboard:

      Label AUC SP-SR MP-SR MPPE (cm)
      tier1 reproduced 0.6052 0.7000 0.7000 14.03
      leaderboard 0.59263 0.66666 0.66666 17.19
      tier2 reproduced 0.4767 0.6000 0.5667 20.39
      leaderboard 0.48510 0.56666 0.53333 19.33
      tier3 reproduced 0.0001 0.0000 0.0000 74.44
      leaderboard 0.00000 0.00000 0.00000 78.10
    • Known failure: an evaluation whose single world turns non-finite stops with ValueError: MPPE requires at least one frame, finite world, and object body, and run.sh reproduce then reports evaluation failed. It is random: it hit tier1 episode 7 once in an earlier full run on this image, the next run of that label succeeded, and our own official evaluation of that episode also succeeded at its second attempt. Re-run the label if it happens.

  • Image neural-motion-v2d-submission:final, image id sha256:c80b68ee1bcbe400c5e585171b553ecd7fdfefb56866454d238acee677c634ad, tarball sha256 e78205652129c789d8e77516df30d77fb23788bef88b12a102ee308d89d0ff3c (10,689,111,366 bytes).
Downloads last month
142