EXYLOS

Robot-ready skill datasets for manipulation policy learning and evaluation.

---

What we do

EXYLOS builds structured datasets for imitation learning, VLA models, policy training, evaluation, and embodiment transfer.

Our datasets span real-world egocentric human capture for motion understanding and retargeting, plus human-seeded robot manipulation workflows performed in simulation. Depending on the release, episodes can include synchronized video and audio, body and hand tracking, robot actions and states, task metadata, outcomes, failure context, object poses, depth, segmentation, and task-specific quality signals.

---

Public sample datasets

This organization hosts compact inspection samples for checking schema, loading data, inspecting trajectories, and evaluating whether the EXYLOS format fits a robotics ML stack.

The two Franka spill-cleanup samples share one task family but serve different inspection needs: the 50-episode release provides broader trajectory coverage without depth, while the 5-episode release exposes the full RGB-D, segmentation, and pose modality stack. The cross-embodiment rich-modality samples add task-matched Panda and WidowXAI trajectories while keeping each embodiment's native schema separate. The egocentric VR sample contains human capture and tracking data rather than robot actions.

---

Dataset format

EXYLOS samples are packaged in LeRobot-compatible or LeRobot-oriented layouts whenever possible. A typical robot-manipulation dataset contains:

README.md
annotations.json
meta/
  info.json
  tasks.jsonl
  episodes.jsonl
  episodes_stats.jsonl
data/
  chunk-000/
    episode_000000.parquet
    episode_000001.parquet
videos/
  chunk-000/
    observation.images.<camera_name>/
      episode_000000.mp4
      episode_000001.mp4

Depending on the dataset, additional external assets may be referenced from parquet rows:

videos/
  chunk-000/
    observation.masks.<camera_name>/
      episode_000000/
        *.png
    observation.depths.<camera_name>/
      episode_000000/
        *.npy

Depth .npy assets are included only in datasets that explicitly advertise depth or RGB-D modalities.

Core signals:

Exact fields vary by dataset, so each repository includes a dataset-specific card.

---

Why EXYLOS datasets are different

• Structured, not raw: episodes include synchronized video, actions, state, metadata, annotations, and quality checks.
• Skill-oriented: each dataset is organized around a manipulation workflow rather than unrelated clips.
• Failure-aware: samples include failed attempts, aborts, collisions, incomplete task executions, and recovery-relevant labels when available.
• Modality-aware: samples range from compact RGB trajectory datasets to richer variants with segmentation masks, depth maps, and object-state streams.
• LeRobot-oriented: data is stored in open formats such as MP4, Parquet, JSON, JSONL, PNG, and NPY.
• Transfer-minded: workflows are captured from human intent in consumer VR and procedurally expanded, with added visual domain randomization for broader policy-learning experiments.

---

Intended use

Public samples are suitable for:

• inspecting EXYLOS schema and annotation conventions
• testing LeRobot-compatible loaders and training pipelines
• running small imitation-learning experiments
• reviewing multi-view video, trajectories, masks, object states, and phase annotations
• evaluating success/failure semantics and cleanup-quality metrics
• comparing task-matched Panda and WidowXAI trajectories without mixing their native control schemas
• inspecting real-world egocentric RGB, audio, body, and hand tracking for retargeting workflows
• assessing whether a custom EXYLOS skill pack would fit your workflow

They are compact inspection datasets, not complete production-scale benchmarks.

---

Commercial datasets and custom skill packs

EXYLOS can generate custom robot-ready skill datasets for pick-and-place, bimanual manipulation, spill cleanup, sorting, binning, object rearrangement, failure recovery, and evaluation/regression sets.

Commercial deliveries can add depth, segmentation masks, object states, event labels, custom cameras, larger episode volumes, stricter QA, and internal-pipeline packaging.

---

License

Licensing is repository-specific. The existing EXYLOS pick-and-place and Franka spill-cleanup samples list Apache 2.0. The egocentric VR sample uses the EXYLOS proprietary evaluation license. The cross-embodiment rich-modality samples do not currently specify a license, so public visibility must not be interpreted as permission to redistribute or reuse them. Always check the individual dataset card and license metadata before using a sample in research, demos, training, or commercial workflows.

---

Contact

• Website: exylos.ai
• Email: contact@exylos.ai
• LinkedIn: Exylos on LinkedIn

If you need structured skill data, send us the target task, robot, modalities, format, evaluation criteria, and timeline.