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

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

Data-driven approaches have emerged as a promising direction for training robotic manipulation policies. Recent robot datasets have grown in scale and diversity by collecting data across multiple embodiments, tasks, and even human demonstration videos. Trained on these increasingly large and diverse manipulation datasets, vision-language-action (VLA) models and world action models (WAMs) generalize across scenes, embodiments, and tasks.

However, these datasets are usually collected in human-absent settings, with the robot acting as the sole agent in the scene. As a result, policies trained on such data are unlikely to perform well in the environments where these robots are meant to be deployed. In homes, factories, and other shared workspaces, a robot must coordinate with co-present humans — following their cues, anticipating their motions, and avoiding collisions. These behaviors are missing from human-absent data not because they are hard to learn, but because they cannot be demonstrated without a human in the scene. For example, a robot cannot learn to hand over a tool if no one is there to receive it, nor to pause for a reaching hand if no hand ever reaches. This gap motivates dedicated datasets that explicitly capture human–robot interaction dynamics and encode collaborative, human-aware behaviors.

We introduce HABIT (Human-Aware Behavior and Interaction Training dataset), a large-scale robot demonstration dataset explicitly designed for human-present environments. In every episode of HABIT, a co-present human shares the workspace with the robot. The dataset comprises 10,563 episodes and 164 hours of bimanual manipulation, spanning 60 tasks. Tasks are organized along three interaction roles that capture distinct dependencies between human and robot: Collaborator, where human and robot jointly accomplish a shared task; Coworker, where human and robot pursue separate tasks within a shared space; and Supervisor, where the human observes and directs the robot. To elicit specific human-aware behaviors such as yielding and gesture-following, we carefully design our collection protocols, while varying other conditions to support generalization.

We verify the effectiveness of HABIT by fine-tuning two open-source VLAs, π0.5 and GR00T N1.6, on a representative six-task subset, and comparing against a matched Robot-only baseline collected without a co-present human. HABIT improves task success rates for both models, with the largest gains on tasks where role-specific coordination is most critical. More notably, training on HABIT gives rise to human-friendly behaviors that emerge directly from data: proactive yielding and collision avoidance under the Coworker role, gesture grounding under Supervisor, and spatiotemporal synchronization under Collaborator. These behaviors reflect the model's internalization of social context when trained on human-present demonstrations. Finally, we show that π0.5 trained on HABIT adapts rapidly to new human-robot interaction tasks.

We believe HABIT is a stepping stone toward robot foundation models that are not merely capable, but genuinely safe and socially compatible in the human-inhabited environments where they will ultimately be deployed.

---

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 — a role taxonomy adopted from the Human-Robot Interaction (HRI) literature, comparable in episode count across roles
🔗 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	60
Episodes	10,563
Frames	5.91 M
Hours	164.19
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 per task (seconds): mean 59.9, median 56.4, range 30.3 – 101.4

Unique subtasks: 157 robot subtasks, 182 human subtasks, 308 unique human–robot subtask pairs.

Per-Role Breakdown

The three roles are comparable in episode count across roles.

Role	Tasks	Episodes	Hours	Robot subtasks / episode	Human subtasks / episode
Collaborator	20	3,198	49.98	3.86	4.70
Coworker	20	3,969	57.33	3.12	4.25
Supervisor	20	3,396	56.88	4.15	3.98

---

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-85.
• 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/configint/HABIT

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

Load with datasets

from datasets import load_dataset

# Small subset covering all 60 tasks (~1 GB) — recommended for quick exploration
ds = load_dataset("configint/HABIT", "sample")

# Full dataset
ds = load_dataset("configint/HABIT", "full")

Configurations

Config	Description	Approx. size
sample	A small subset covering all 60 tasks (1 episode per task) for quick exploration of the schema and structure	~1 GB
full	The complete dataset: 10,563 episodes, 164.19 hours, 60 tasks	—

---

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, 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 60 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).

---

Citation

@article{song2026habit,
  author  = {Jaehwi Song and Suchae Jeong and Byeongguk Jeon and Sungdong Kim and Minjoon Seo and Hyungmok Son and Kimin Lee},
  title   = {HABIT: Human-Aware Behavior and Interaction Training Dataset for Robot Manipulation},
  journal = {arXiv preprint arXiv:2606.31682},
  year    = {2026}
}