Simaihub Mobile Robot Navigation Training Dataset for research

Data Product Features

• episode_id — Unique episode identifier within the package.
• step_id / timestamp — Time-step index (and optional timestamp) within an episode.
• observations — Multi-modal robot observations as configured (e.g. LiDAR / vision / proprioception / goal-relative features).
• actions — Continuous control commands (typically linear and angular velocity).
• rewards — Scalar reward and/or decomposed reward terms for offline RL.
• dones / truncated — Episode termination and truncation flags.
• scene / route metadata — Scene name, route type, task labels (e.g. straight, avoid, recovery).
• success / failure tags — Labels distinguishing successful episodes from curated failure or recovery samples (when included).

Distribution

• Primary format: HDF5 (training package); documentation in Markdown / JSON schema; optional preview tables.
• Structure: Episode-grouped trajectories with per-step obs / action / reward / terminal fields.
• Data Volume: Multi-scene package across Hospital, Office, Warehouse, Simple Room; exact episode and step counts are stated on the release datasheet for each version.
• Delivery: Licensed download or private enterprise delivery; sample subset may be published for evaluation only under the same EULA.

Usage

• Imitation learning / behavior cloning: Train navigation policies from expert demonstrations.
• Offline reinforcement learning: Train or evaluate policies from logged (s, a, r, s', done) trajectories.
• Policy pre-training: Warm-start navigation agents before online fine-tuning in sim or on hardware.
• Generalization studies: Cross-scene evaluation and curriculum design using route / lighting / obstacle diversity.
• Sim-to-real research: Domain-randomized simulation data as a source domain for transfer experiments (subject to Licensee’s own validation on real robots).

Coverage

• Geographic Coverage: Simulation-only indoor environments (not real-world geo-tagged logs). Select any two scenes from the following list: Hospital, Office, Warehouse, Simple Room.
• Time Range: Synthetic collection window per release version (see CHANGELOG / datasheet for the shipping build).
• Demographics (if applicable): N/A — robotics simulation trajectories; no human personal data intended.
• Robot / task scope: Differential-drive / AMR-style indoor point-to-point navigation with multi-pose routes and recovery behaviors.

License

AI Training Rights

• Access and use the Dataset for academic research and/or internal commercial research and product development within Licensee’s organization.
• Reproduce and internally modify the Dataset to train, evaluate, or fine-tune machine-learning or robotics models.
• Deploy resulting models inside Licensee’s own products or services (e.g. robots, AMRs, edge devices, customer-facing software), provided Licensee does not redistribute the Dataset or any substantial extract/reconstruction of it.
• Publish research results that reference the Dataset, subject to attribution (Section 4).

• The Dataset itself may not be sold, redistributed, shared, or posted (public or private platforms, public cloud links, GitHub, etc.) outside licensed internal use.
• No creation/publication of derivative datasets for others without a separate written agreement.
• Without a separate Commercial License Agreement, Licensee may not sell or distribute a trained model as a standalone dataset substitute or downloadable “navigation foundation package” whose primary value is replacing this Dataset.
• No competitive use that builds a product/service competing with Licensor’s offerings without written consent (Section 3.4).
• No removal of copyright notices, watermarks, or license identifiers.
• Licensee must comply with applicable laws (including export control and data-protection rules where relevant).

Who Can Use It

• Robotics / embodied-AI researchers: Academic papers, benchmarks, and non-commercial experiments (with attribution).

Data Dictionary

| Column Name | Data Type | Description | Possible Values/Notes |
|-------------|-----------|-------------|-----------------------|
| episode_id | string / int | Episode identifier | Unique per episode in package |
| step_id | int | Step index within episode | 0 … T-1 |
| observations | nested array / group | Sensor and state features | Schema versioned per release; see datasheet |
| actions | float32[2] (typical) | Control command | e.g. [v_linear, ω_angular] |
| rewards | float32 / struct | Reward signal | Scalar and/or decomposed terms |
| dones | bool | Episode terminated | true/false |
| truncated | bool | Episode truncated (timeout etc.) | true/false |
| scene | string | Simulation scene | hospital, office, warehouse, simple_room |
| route_type | string | Route / task class | straight, avoid, mixed, recovery, … |
| outcome | string | Episode outcome label | success, failure, recovery (if provided) |

• Sample vs full pack: Public samples (e.g. on Hugging Face) are for evaluation; full commercial packages are delivered under this EULA.
• Pricing SKUs (Research / Commercial Standard / Enterprise) are commercial packaging; legal permissions are defined only by LICENSE.md and any written addendum. Do not rely on marketing pages alone.
• Provenance: Current product generation stack is NVIDIA Isaac Sim + Isaac Lab. Listing materials and LICENSE attribution/third-party sections must stay consistent with the shipping build (see license analysis below for required updates).