video video 5 59.6 | label class label 50
classes |
|---|---|
0adjust_bottle | |
0adjust_bottle | |
0adjust_bottle | |
0adjust_bottle | |
0adjust_bottle | |
0adjust_bottle | |
0adjust_bottle | |
0adjust_bottle | |
0adjust_bottle | |
0adjust_bottle | |
1beat_block_hammer | |
1beat_block_hammer | |
1beat_block_hammer | |
1beat_block_hammer | |
1beat_block_hammer | |
1beat_block_hammer | |
1beat_block_hammer | |
1beat_block_hammer | |
1beat_block_hammer | |
1beat_block_hammer | |
2blocks_ranking_rgb | |
2blocks_ranking_rgb | |
2blocks_ranking_rgb | |
2blocks_ranking_rgb | |
2blocks_ranking_rgb | |
2blocks_ranking_rgb | |
2blocks_ranking_rgb | |
2blocks_ranking_rgb | |
2blocks_ranking_rgb | |
2blocks_ranking_rgb | |
3blocks_ranking_size | |
3blocks_ranking_size | |
3blocks_ranking_size | |
3blocks_ranking_size | |
3blocks_ranking_size | |
3blocks_ranking_size | |
3blocks_ranking_size | |
3blocks_ranking_size | |
3blocks_ranking_size | |
3blocks_ranking_size | |
4click_alarmclock | |
4click_alarmclock | |
4click_alarmclock | |
4click_alarmclock | |
4click_alarmclock | |
4click_alarmclock | |
4click_alarmclock | |
4click_alarmclock | |
4click_alarmclock | |
4click_alarmclock | |
5click_bell | |
5click_bell | |
5click_bell | |
5click_bell | |
5click_bell | |
5click_bell | |
5click_bell | |
5click_bell | |
5click_bell | |
5click_bell | |
6dump_bin_bigbin | |
6dump_bin_bigbin | |
6dump_bin_bigbin | |
6dump_bin_bigbin | |
6dump_bin_bigbin | |
6dump_bin_bigbin | |
6dump_bin_bigbin | |
6dump_bin_bigbin | |
6dump_bin_bigbin | |
6dump_bin_bigbin | |
7grab_roller | |
7grab_roller | |
7grab_roller | |
7grab_roller | |
7grab_roller | |
7grab_roller | |
7grab_roller | |
7grab_roller | |
7grab_roller | |
7grab_roller | |
8handover_block | |
8handover_block | |
8handover_block | |
8handover_block | |
8handover_block | |
8handover_block | |
8handover_block | |
8handover_block | |
8handover_block | |
8handover_block | |
9handover_mic | |
9handover_mic | |
9handover_mic | |
9handover_mic | |
9handover_mic | |
9handover_mic | |
9handover_mic | |
9handover_mic | |
9handover_mic | |
9handover_mic |
IOI Baseline: RoboTwin2.0 Video Generations
1,000 full-episode, prediction-only videos from two action-conditioned world models, using their 60,000-step raw-model checkpoints:
IRASim/: 500 videos.Ctrl-World/: 500 videos of the same source episodes.episodes.json: exact source episode mapping, split, frame ranges, seeds, model checkpoint fingerprints, output paths and SHA-256 checksums.
Source Episodes
Source: official RoboTwin2.0 dataset,
revision 3dc3b798668feb99ac61cc9086d84cbcc3d79186.
Only dataset/<task>/demo_clean.zip, embodiment aloha_agilex, is used.
The upstream dataset declares an MIT license; consult its dataset card for terms.
The selection covers all 50 clean tasks, with ten episodes per task:
five held-out validation episodes and five training episodes. These are local
experimental splits, not an official upstream split. The training examples are
seen-data diagnostics; do not pool them with held-out results or describe all
500 episodes as validation. Selection seed: 20261028; no quality-based pruning.
For each selected episode, episodes.json specifies both the source ZIP path
and the exact HDF5 member path inside it, for example:
dataset/open_microwave/demo_clean.zip
open_microwave/aloha_agilex/data/episode_0000016.hdf5
Use vision/cam_head/colors for corresponding GT frames. All source frame
indices from zero through frames - 1 are retained at 15 FPS. Original GT
videos, raw HDF5 data, model weights and credentials are not included here.
Generation Settings
Head camera only, RGB, 320 x 256 pixels, 15 FPS. Source frames were resized directly with bilinear interpolation, not letterboxed or center-cropped. Both models use 14D absolute joint positions: six left-arm joints, left gripper, six right-arm joints, right gripper. Joint normalization uses frozen training statistics; grippers use a fixed opening mapping.
Only the first real image conditions the visual rollout. Later images are generated autoregressively using recorded joint trajectories. These are action-conditioned video rollouts, not closed-loop robot policy evaluations. The initial conditioned-frame reconstruction is included in each video.
- IRASim: 50 denoising steps, guidance 1, current generated frame plus 15 new frames per chunk; raw model weights, not EMA.
- Ctrl-World: 25 denoising steps, guidance 1, six consecutive generated history frames plus the current frame and four future frames. Initial history is left-padded with the initial frame. This contiguous-history adaptation is not a claim of reproducing the official sparse-history recipe exactly.
Code references: IRASim and Ctrl-World. These are adapted-model outputs, not official benchmark results from those projects.
Video Export and Verification
The published files contain only prediction pixels, without GT panels, comparison labels or audio. They are extracted from the right-hand image of the saved comparison videos (crop x=320, y=32, width=320, height=256), then re-encoded as H.264, yuv420p, CRF 18. Thus they include an additional lossy encoding generation and are not raw decoder tensors. Inference was not rerun for this export. For pixel-exact model evaluation, use the original generated frames/latents rather than these presentation MP4s.
Every exported video was fully decoded and checked for its expected frame
count, resolution and FPS. File sizes and SHA-256 checksums appear in
episodes.json. Generated videos can exhibit visual drift, incorrect contact
or incorrect motion; inclusion does not establish successful task execution.
- Downloads last month
- 280