--- license: mit pretty_name: Future Sight PushT task_categories: - robotics tags: - robotics - world-modeling - pusht - tfrecord - video --- # Future Sight PushT Future Sight PushT is a wind tunnel for studying how training-data scale and composition affect world-model learning and generalization. It pairs large passively collected corpora with smaller, iterative active-exploration corpora in a common PushT environment. This controlled setting helps isolate those effects from the scale and potential contamination of large, heterogeneous datasets. ## Collection preview ### Passive collection Expert PPO, perturbation and recovery, random walk, and random actions: four ways to collect large training corpora with different behavior distributions. ### Active exploration AIM, DADS, DIAYN, ICM, and METRA: smaller collections produced by task-specific exploration methods. The selected AIM episode is from iteration 1; DADS, DIAYN, ICM, and METRA are from iteration 4. Both videos show selected recorded trajectories, up to 100 frames each. Playback is 10 display frames/s; short episodes repeat, with the frame counter marking each restart. This display rate does not specify collection frequency. These clips illustrate individual trajectories, not aggregate performance or coverage. Still figures at the 5-second display point: [passive PNG](assets/passive-figure.png) · [PDF](assets/passive-figure.pdf); [active PNG](assets/active-figure.png) · [PDF](assets/active-figure.pdf). ## Dataset at a glance | | Passive training | Active training | Evaluation | | --- | ---: | ---: | ---: | | Episodes | ≈1.21 million | 17,000 | 400 | | Compressed size | 486.249 GB | 8.901 GB | 0.239 GB | Active episode counts include all five collection iterations and both AIM ablations. Each evaluation collection contains 100 episodes. Passive counts are rounded; all sizes are compressed decimal GB. The previewed training corpora use 144×144 RGB PNG observations and six-dimensional continuous actions and a single camera view. Trajectories are stored as TFRecords, with JSON corpus metadata and CSV per-trajectory metadata. Evaluation records use the same decoded RGB/action dimensions and additionally include PNG segmentation. ## Collection guide Passive collections represent data that can be gathered at scale: | Collection | Behavior | Episodes | Compressed size | | --- | --- | ---: | ---: | | Expert PPO | Task-directed expert trajectories | ≈307,000 | 56.976 GB | | Recovery | Policy rollouts with perturbations and recovery | ≈300,000 | 118.956 GB | | Random walk | Temporally correlated random actions | ≈300,000 | 156.312 GB | | Uniform random | Uniform random actions | ≈300,000 | 154.005 GB | Active collections represent smaller, task-specific exploration rounds. Each method has five corpora, `iter-0` through `iter-4`, and each iteration has eight shards. | Method | Role | Episodes, all five iterations | Compressed size | | --- | --- | ---: | ---: | | AIM | Action Influence Maximization | 5,000 | 2.667 GB | | AIM no-MI | AIM mutual-information ablation | 5,000 | 2.617 GB | | AIM no-SN | AIM spectral-normalization ablation | 5,000 | 2.631 GB | | DADS | Dynamics-Aware Discovery of Skills | 500 | 0.260 GB | | DIAYN | Diversity Is All You Need | 500 | 0.223 GB | | ICM | Intrinsic Curiosity Module | 500 | 0.244 GB | | METRA | Metric-aware abstraction | 500 | 0.260 GB | Evaluation data is separate from training data: | Corpus | Behavior | Episodes | | --- | --- | ---: | | BC | Behavior-cloned policy rollouts | 100 | | BC failures | Alternating expert and push-away controller segments | 100 | | PushAway | Push-away controller rollouts | 100 | | Interaction | Human teleoperation with simulator segmentation masks | 100 | ## Files and download ```text passive-collection//train/shard_N.tar.zst active-exploration//iter-N/shard_N.tar.zst evaluation//shard_0.tar.zst manifest.json ``` [`manifest.json`](manifest.json) records every archive path, byte size, and SHA-256 checksum. Sizes above are compressed decimal GB; extraction requires additional disk space. Install the [Hugging Face CLI](https://huggingface.co/docs/huggingface_hub/guides/cli), then download only the data needed for an experiment. For one archive: ```sh hf download smbml/future-sight-pusht evaluation/bc/shard_0.tar.zst \ --repo-type dataset --local-dir ./future-sight-pusht ``` For one complete active-exploration iteration: ```sh hf download smbml/future-sight-pusht \ --repo-type dataset \ --include "active-exploration/aim/iter-0/*.tar.zst" \ --local-dir ./future-sight-pusht ``` Add `--revision FULL_COMMIT_HASH` to pin a reproducible snapshot. Download `manifest.json` at the same revision when verifying sizes or checksums. ## Extract and read With GNU tar and `zstd` installed: ```sh mkdir -p ./future-sight-data tar --zstd -xf ./future-sight-pusht/evaluation/bc/shard_0.tar.zst \ -C ./future-sight-data ``` This creates `future-sight-data/pusht_bc_1pct_n100/` with TFRecords and metadata. Archive paths organize downloads; extracted top-level directory names reflect the corpus stored in each archive. For sharded training corpora, extract each shard into a separate directory so shard-local metadata files do not overwrite one another. Read the TFRecords and metadata from all eight directories to use a complete iteration or passive corpus. ## License MIT. ## Citation If you use this dataset, please cite: ```bibtex @inproceedings{bateman2026training, title = {Training Controllable World Models via Action Influence Maximization}, author = {Bateman, Samuel M. and Yin, Tenny and Zheng, Chongyi and Huang, Lei and Wang, Brian and Eysenbach, Ben and Fern{\'a}ndez Fisac, Jaime and Shah, Dhruv}, booktitle = {Conference on Robot Learning (CoRL)}, year = {2026} } ``` *Paper link coming soon.*