HABIT: Human-Aware Behavior and Interaction Training Dataset for Robot Manipulation

⚠️ Anonymous release. Authors and institutional information are intentionally withheld. This dataset card will be updated when these details become available.

TL;DR: HABIT is a large-scale robot demonstration dataset for human-present environments, designed to teach robot policies human-aware behaviors.

Keywords: Robot Manipulation Dataset, Human-Robot Interaction, Vision-Language-Action Model

---

Overview

Recent progress in general-purpose robot policies has been driven by large-scale demonstration datasets. However, existing datasets are collected almost exclusively in human-absent settings, with the robot acting as the sole agent in the workspace. As a result, policies trained on such data may execute tasks competently in isolation but cannot acquire human-aware behaviors such as yielding to a reaching hand, handing over an object, or following a pointing gesture. These behaviors are not merely hard to learn. They are structurally absent from the data, because they cannot be demonstrated without a human in the scene.

HABIT addresses this gap by providing a large-scale robot demonstration dataset where every episode features a co-present human sharing the workspace with the robot. Tasks are organized along three interaction roles that capture distinct human-robot interaction, with rich subtask-level annotations on both the human and the robot side.

Fine-tuning open-source VLAs on HABIT elicits human-aware behaviors that robot-only training fails to produce: yielding under Coworker, gesture grounding under Supervisor, and spatiotemporal synchronization under Collaborator. Mid-training on HABIT further enables sample-efficient adaptation to new collaboration tasks.

By introducing human presence as a new axis of dataset diversity, HABIT extends robot policies to environments shared with humans.

---

Highlights

🤝 Human-present demonstrations	Every episode contains a co-present human partner, captured under a reactive collection protocol
🎭 Three interaction roles	Collaborator, Coworker, and Supervisor — balanced in episode count by design
🔗 Task Workflow	Graph-structured task representation explicitly capturing cross-agent dependencies
👫 Subtask-level annotations	Recorded subtask boundaries for both human and robot, captured during the demonstration
📷 Five RGB camera streams	3 robot-side + 2 human-side viewpoints capture thorough human-robot interaction
🦾 Bimanual manipulation	Two Franka Research 3 (FR3) arms on a shared frame

---

Dataset Statistics

Aggregate Scale

Tasks	48
Episodes	7,518
Frames	4.03 M
Hours	111.83
Roles	3 (Collaborator / Coworker / Supervisor)
Cameras per episode	5 (3 robot-side + 2 human-side)
Robot platform	Bimanual Franka Research 3
Format	LeRobot v2.0

Episode length (seconds): mean 53.6, median 49.3, range 17.1 – 150.5

Unique subtasks: 129 robot subtasks, 144 human subtasks, 256 unique human–robot subtask pairs.

Per-Role Breakdown

The three roles are balanced in episode count by design.

Role	Tasks	Episodes	Hours	Robot subtasks / episode	Human subtasks / episode
Collaborator	16	2,521	36.37	3.98	4.64
Coworker	15	2,518	35.84	2.99	4.18
Supervisor	17	2,479	39.62	3.82	3.61

---

Three Interaction Roles

Collaborator

The human and robot jointly accomplish a shared goal through direct physical interaction (e.g., handing over an object, jointly holding a bucket). The robot must coordinate spatially and temporally with the human.

Coworker

The human and robot share a goal and a workspace, but without direct physical contact — each agent independently handles its own portion of the task. The robot must avoid collisions with the human to ensure safety.

Supervisor

The human directs the robot through explicit cues such as gestures or behavioral instructions, and the robot must perceive the human's intent from visual input alone.

---

Hardware Setup

• Robot: Bimanual Franka Research 3 with Robotiq 2F-25.
• Workspace: A front table placed between the human and the robot (directly shared workspace) and a side table located beside the human (dedicated to human-side activities).
• Cameras (5 × RGB):
  • 1× robot-center — angled forward to capture both the human and the shared workspace
  • 2× wrist-mounted — one per arm
  • 1× human head-mounted (egocentric) — captures the human's first-person perspective; useful for grounding pointing gestures whose referents are difficult to identify from robot-side views alone
  • 1× exocentric — observes the entire human-robot workspace, providing a holistic view of the interaction
• Teleoperation: Meta Quest 3 controllers, building on the DROID codebase. Action representations are recorded in joint-space, Cartesian-space, and gripper-state form, so downstream users may train policies in whichever action space their model expects.

---

Data Collection Protocol

Each episode is recorded by two operators acting in distinct roles — a robot operator who teleoperates the bimanual FR3, and a human operator who physically performs the human-side activity. Both press their own foot pedal to mark each subtask boundary online (see Subtask Annotations).

The protocol is shaped by three principles:

1. Reactive interaction. Each operator acts only after directly observing the partner. Any coordination outside the recorded streams (rehearsal, eye signals, verbal cues) is prohibited, so every demonstrated action is grounded in cues that are also visible to the policy.
2. Targeted behavior elicitation. Three design choices make specific human-aware behaviors appear in the data: yielding (robot retracts when about to collide with the human), temporal adaptation (human operator's speed is varied across episodes), and gesture grounding (human varies the wait time before pointing).
3. Diversification. Within each task, clothing color and object-interaction order are varied across episodes. Across tasks, the dataset spans multiple human operators with different body types.

---

Get Started

Download

# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install

# Clone the full dataset
git clone https://huggingface.co/datasets/habit-anonymous/HABIT

# Or skip large files and fetch on demand
GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/datasets/habit-anonymous/HABIT

Load with datasets

from datasets import load_dataset

# Small subset covering all 48 tasks (~1 GB) — recommended for quick exploration
ds = load_dataset("habit-anonymous/HABIT", "sample")

# Full dataset (~160 GB)
ds = load_dataset("habit-anonymous/HABIT", "full")

Configurations

Config	Description	Approx. size
sample	A small subset covering all 48 tasks (1 episodes per task) for quick exploration of the schema and structure	~1 GB
full	The complete dataset: 7,518 episodes, 111.83 hours, 48 tasks	~160 GB

---

Dataset Structure

LeRobot version reference: v2.0

Each configuration (full and sample) is laid out independently under its own top-level folder:

{full,sample}/
├── meta/
│   ├── episodes.jsonl            # per-episode metadata (index, task, length)
│   ├── tasks.jsonl               # list of high level language instructions
│   ├── subtasks.jsonl            # list of robot low level language instructions
│   ├── human_subtasks.jsonl      # list of human low level language instructions
│   ├── info.json                 # schema, fps, features, path templates
│   ├── modality.json             # state and action modalities for policy training
│   └── stats.json                # per-feature normalization statistics
├── data/
│   └── chunk-XXX/
│       └── episode_XXXXXX.parquet
└── videos/
    └── chunk-XXX/
        └── observation.images.<view>/
            └── episode_XXXXXX.mp4

Storage rules

• Each Parquet file stores one complete episode (state, action, timestamps, language instruction (high level, robot low level, human low level)).
• Each MP4 file stores one camera stream for one episode.
• The dataset schema is primarily defined in meta/info.json.

---

Features Schema

The full per-step state and action schema — including all available modalities (Cartesian / joint / gripper, target / delta / accumulated, etc.) and their slicing into the underlying parquet columns — is defined in meta/modality.json. Refer to that file for the authoritative key list and dimensions.

Subtask Annotations

A distinguishing feature of HABIT is that subtask completion timestamps are recorded during the demonstration itself via foot pedals — one pedal pressed by the robot operator upon completing each robot subtask, and a separate pedal pressed by the human operator upon completing each human subtask. This provides fine-grained alignment between human and robot subtask boundaries that is not available in conventional teleoperation pipelines.

Concretely, every step in each episode's Parquet file carries the currently active subtask index for both agents. These are HABIT-specific fields added on top of the standard LeRobot v2.0 schema:

Column	Agent	Maps to
low_level_task_index	Robot	task_index in meta/subtasks.jsonl (robot low-level instructions)
human_role_subtask_index	Human	task_index in meta/human_subtasks.jsonl (human low-level instructions)

Because both indices are aligned to the same step, the per-step pair (low_level_task_index, human_role_subtask_index) directly captures which human and robot subtasks are co-occurring at any moment of an episode.

Camera Feature Keys

Camera streams are stored under observation.images.<view> keys following LeRobot convention. The dataset includes five RGB streams per episode:

Stream	Side	Purpose
observation.images.front_view	Robot	Forward-facing center view; both human and shared workspace
observation.images.left_wrist_view	Robot	Wrist-mounted, left arm
observation.images.right_wrist_view	Robot	Wrist-mounted, right arm
observation.images.human_front_view	Human	Head-mounted, human first-person view
observation.images.exo_view	External	Holistic exocentric view

💡 Robot-only training. For VLA fine-tuning experiments in our paper, only the 3 robot-side cameras (front_view, left_wrist_view, right_wrist_view) are used. The two human-side streams are provided for future work (e.g., grounding pointing gestures, studying human-side behavior).

---

Tasks

The dataset spans 48 tasks across the three roles. The full task list (high-level instructions) is available in meta/tasks.jsonl.

Each episode in meta/episodes.jsonl carries a sid (scenario ID) field that links the episode to its detailed task documentation. For each sid, we provide a complete description in task_details/ covering:

• Environment setup — initial object placements and workspace configuration
• High-level instruction — the overall goal given to both operators
• Low-level instructions — per-agent (human / robot) subtask sequences
• Workflow graph — a directed graph encoding the precedence relations between human and robot subtasks (image embedded inline; raw PNGs under figures/workflows/)

Files are named by sid (e.g., S7025.md). For a few tasks, the same underlying task was split across multiple sids purely to introduce execution variations (e.g., different object placements or interaction orders). These share a single description, and the file name joins all member IDs (e.g., S8102_S8103_S8097.md).

Below is the representative six-task subset used for evaluation in the paper:

Collaborator (spatiotemporal coordination)

• Table Serving — Robot lifts dishware from the tray the human approaches; human lays a napkin underneath
• Shelf Cleaning — Tier-by-tier cleaning, bracketed by duster handovers between human and robot

Coworker (reactive collision avoidance)

• Waste Sorting — Robot sorts cans, human sorts bottles and plastics, into shared bins (moderate workspace overlap)
• Box Packaging — Each agent packs their own box from a shared item pool (high workspace overlap)

Supervisor (gesture grounding)

• Donut Serving — Robot picks the donut indicated by human pointing (single-stage trajectory)
• Food Storage — Robot places bread in the container indicated by human pointing (two-stage trajectory)

---

Intended Uses

• Fine-tuning vision-language-action (VLA) models and other robot policies for human-aware bimanual manipulation
• Studying human-robot collaboration, coordination, and gesture grounding
• Mid-training as a prior for downstream human-robot collaboration tasks (sample-efficient adaptation)
• Benchmarking out-of-distribution robustness to human-centric distribution shifts (clothing color, body silhouette)

---

Limitations

• Single embodiment. All demonstrations are on the bimanual Franka FR3 morphology. No cross-embodiment data.
• Limited collector pool. On the order of 10 human collectors total — insufficient to fully cover the diversity of body silhouettes, motion styles, and clothing patterns a deployed robot would encounter.
• One-to-one setting. Each episode contains exactly one human and one robot. Multi-human and multi-robot scenarios are out of scope.
• Single laboratory environment. All data are collected in one lab. Cross-environment generalization is not directly supported by this dataset.

---

Ethics Statement

• Informed consent. All human collectors participated voluntarily and provided written informed consent, prior to data collection, for the recording, processing, and public release of video data containing their image — including their unblurred face. Collectors were also informed that they may withdraw their consent at any time and request removal of their data from the release.
• Institutional review. The data collection protocol — including consent procedures, compensation, and downstream data handling — was reviewed and approved by an internal ethics review, as well as a separate internal privacy / personal-data-handling review. Specific institutional references are withheld in this anonymous release.
• Privacy protection. Although collectors consented to release with unblurred faces, we adopted a more conservative privacy posture and applied automatic face blurring to all released video streams. Despite this, residual identifying cues (e.g., body silhouette, clothing, hands) may remain in the data, and a small number of frames may not be fully de-identified due to detector failures. Users should treat the dataset accordingly and refrain from attempts to re-identify individual collectors.
• Scope of release. Audio was never recorded during data collection. Only video and robot-state streams are released, and no names or other personally identifying metadata are distributed alongside the dataset.

---

License

Released under CC BY 4.0. You are free to share and adapt the data for any purpose, including commercially, with appropriate attribution.

---

Citation

@misc{habit2026,
  title        = {HABIT: Human-Aware Behavior and Interaction Training Dataset for Robot Manipulation},
  author       = {Anonymous},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/habit-anonymous/HABIT}}
}

The citation will be updated once author information can be disclosed.

---

Changelog

• v0.1 — Initial anonymous release.