Habitat 3.0

Habitat 3.0: A Co-Habitat for Humans, Avatars and Robots

How to read this picture

Habitat 3.0 is a simulator release with two tasks in which a simulated Spot robot works with simulated people. The robot either finds and follows a person or tidies a home together with one.123

Sources
Last checked 10 Oct 2026Full entry49 of 58 facts checked at the sourceNext check 8 Apr 2027
Runs in
Simulation2
Checked against real robots
Not checked
Skill
Working with people
Robot
Arm on wheels2
Used by
A few papers report on the tasks. Most citing papers use it as a simulator.456+1
342 citations
Licence
MIT89
Commercial use: not allowed

What a score here does not tell you Inferred

  1. How well a policy (the robot's control model) will do on a real robot.We found no study that compares Habitat 3.0 results with real robots.
  2. How well a robot will work with real people.Only one study has tested this. It had 30 people and two policies.
  3. How a robot copes with natural human behaviour.The simulated people can only walk and reach.

Comparisons with real robots

Not checked Inferred2105+1

Details

The paper has no real-robot experiments; Meta's blog lists deploying the learned models in the physical world as a next step. The blog's '20% success rate in the physical world' refers to the HomeRobot OVMM benchmark, not these tasks.

Our assessment Opinion

Scores show how well a policy works with simulated people. Nothing links them to real robots.

Reasoning

Habitat 3.0 scores measure how well a policy works with scripted or learned simulated people in simulated homes. The human study suggests the order of two policies carries over to real people in simulation. Nothing links the scores to real robots.

Confidence: medium

Treat the few published numbers as reference points.

Reasoning

Few papers report on these exact tasks and the evaluation set varies, so published numbers work as reference points. They do not form a leaderboard.

Confidence: medium

It is mostly used as the base for newer benchmarks.

Reasoning

Habitat 3.0 matters most as the base for later benchmarks such as PARTNR and Social-HM3D. With Meta's maintenance ended in 2026, new work will depend on forks.

Confidence: medium

Known problems 5

  1. There is no fixed evaluation set

    Sources use 150, 400, 500 or 1,200 evaluation episodes.2116

    Details

    The paper's appendix lists 150 test episodes (15 in each of 10 test scenes) and 1,200 validation episodes. The habitat-baselines README says the paper's numbers come from 'the full evaluation dataset (1200 episodes)' and gives a command that evaluates 500 episodes. Its released social navigation checkpoint uses Spot's body stereo depth in place of the paper's humanoid GPS, and its expected output (found-human rate 0.9020) is not in the paper's metric format. A 2026 paper evaluates on 400 episodes over 12 environments.

  2. The project is unmaintained and released files have open bugs

    Meta has ended maintenance, and reported bugs in the data remain open.121314+1

    Details

    Meta stopped maintaining habitat-lab and habitat-sim after v0.3.4 (2026-05-07). Open issues report errors when evaluating the released social navigation checkpoint (issue #1839, 2024-03) or the social navigation baseline (issue #1684, 2023-11), and objects missing from the provided episodes (issue #2105, 2024-11). Users posted workarounds; one says the fix still fails on some scenes.

  3. Collision rate penalises following for longer

    A robot that follows the person for longer has more chances to collide. So the following rate and the collision rate trade off against each other.52

    Details

    CR is the share of episodes that end in a collision, and episodes run to 1,500 steps instead of stopping once the robot has followed long enough. The SDA authors note that a robot which finds and follows the person longer has more chances to collide. When they ended episodes after 400 following steps, SDA and the baseline had similar collision rates (0.39 and 0.38). In the original table even the privileged heuristic expert ends 52% of episodes in a collision.

  4. Simulated people behave simply

    The simulated partners act less like real people than the scores imply.2

    Details

    The simulated people only walk and reach; in Social Navigation the person walks shortest paths to random points. In the human study, real partners adapted to the robot and reached success 1 in all episodes, while simulated partners left success rates of 48.52% to 71.79% with unseen partners. The authors conclude their simulated partners do not accurately capture human-robot dynamics.

  5. Text and tables disagree in places

    Several numbers in the text differ from the paper's tables.2

    Details

    The text gives 71.7% for the best unseen-partner success, the table 71.79. The text says Plan-Pop4's RE drops from 105.49 to 101.99, but 101.99 is Plan-Pop3's value in the table; Plan-Pop4's is 103.53. The text says learned skills drop success to 41.96%, the table shows 41.09. The main text gives the humanoid 188±2 FPS in one environment; Appendix F.1 gives 155±26.

Details

About

What it is
Benchmark32
More

The project page presents the two tasks 'aiming at reproducible and standardized benchmarking'. The paper calls Habitat 3.0 a simulation platform.

Built by
Meta FAIR210
More

23 authors. The paper's footnote says 'Work done at Fair, Meta'; no other affiliations are given. The launch blog is a Meta AI post.

Released
October 2023, at ICLR 202411610+1
More

arXiv v1 on 2023-10-19. habitat-sim v0.3.0 (2023-10-19) and habitat-lab v0.3.0 (2023-10-20) shipped it; the Meta blog post followed on 2023-10-20.

Version
habitat-lab v0.3.4 and habitat-sim v0.3.3161718
More

v0.3.0 · 2023-10-19 (sim) and 2023-10-20 (lab): first release with Habitat 3.0.1617

v0.3.1 to v0.3.3 · habitat-lab 2024-03-15, 2024-10-30, 2025-01-27; habitat-sim 2024-03-15, 2024-10-30, 2026-02-12.1617

habitat-lab v0.3.4 · 2026-05-07. Last release maintained by Meta.1612

hab3_episodes · Episode data for both tasks plus a social navigation checkpoint. Last modified 2023-11-28.18

Last update
May 2026. This was Meta's final release.161219+1
More

habitat-lab v0.3.4 on 2026-05-07. The same day a README notice said Meta no longer develops or maintains the project beyond v0.3.4. The task episodes last changed on 2023-11-28.

The habitat-sim README carries the same notice. habitat-sim's last release is v0.3.3 (2026-02-12).

Status
Meta stopped maintaining it in May 2026. Inferred122018+1
More

Meta stopped development after habitat-lab v0.3.4 (2026-05-07). The task data has not changed since 2023-11.

By the taxonomy's one-year rule the platform was updated recently, but the maintainer has announced no further work, so we use 'dormant'. Forks may continue.

Setup

Runs in
Simulation2
More

All experiments, including the human-in-the-loop study, run in simulation.

Simulator
Habitat-Sim and Habitat-Lab v0.323
More

Habitat-Sim and Habitat-Lab v0.3. Simulated people use SMPL-X bodies with walking and reaching motions. The paper reports 1190 FPS with a humanoid and a robot (1345 FPS with two robots).

The intro gives 1190 FPS. Appendix F.1 reports 1,100 to 2,290 FPS with 16 environments on one V100 GPU.

Robot
Arm on wheels2
More

The basic entry is kept. The partner is a simulated person, not a robot, so 'humanoid' is not added.

Robot model
Spot (simulated)221
More

The Spot model (URDF) is 'provided courtesy of Boston Dynamics, all rights reserved', redistributed with written permission (dataset card).

Setting
Whole home2
Tasks
2 tasks23
More

2 tasks: Social Navigation (find a person and follow at 1 to 2 m) and Social Rearrangement (move two objects to goals together with a person).

Scenes
59 homes: 37 for training, 12 for validation and 10 for testing210
More

59 HSSD homes: 37 train, 12 validation and 10 test, taken from the 211 scenes of HSSD-200

Changes at test
Homes and partners not seen in training2
More

Evaluation uses homes and object placements not seen in training. Social Rearrangement also tests partners not seen in training (zero-shot coordination), which has no taxonomy value.

Scoring and access

Scored by
Success rate, Path efficiency2
More

Following rate, collision rate and relative efficiency have no taxonomy value.

Score
Rates of finding, following and collision, and success at the joint task2
More

Social Navigation: finding success S, SPS (S weighted by path steps against an oracle that knows the person's route), following rate F (share of possible steps spent 1 to 2 m away facing the person) and collision rate CR (share of episodes ending in a collision). Social Rearrangement: success rate SR (both objects placed) and relative efficiency RE (steps for the person alone compared with the team), tested with the training partners and with 10 partly unseen partners.

SDA (2024) adds episode success ES: find the person and follow for 400 steps at a safe distance.

Trials
1,500-step episodes, with 3 training seeds2116
More

Social Navigation episodes · 1,500 steps, ending early on a collision. The robot starts at least 5 m from the person. Start positions and the person's path are fixed across baselines.2

Seeds · Every baseline is trained with 3 random seeds and results are averaged.2

Episode counts differ by source · Paper appendix: 15 episodes in each of 10 test scenes (150) and 100 in each of 12 validation scenes (1,200). habitat-baselines README: the paper used 'the full evaluation dataset (1200 episodes)' and its own command evaluates 500. CommNav (2026): 400 test episodes over 12 environments.21118+1

Who runs it
Each team tests its own model Inferred312
More

No organiser runs submissions. Each paper runs its own evaluation.

Error bars
Usually reported Inferred25
More

The paper reports means with ± over 3 seeds and 95% confidence intervals for the human study. SDA reports ± too. We checked these two papers only.

Leaderboard
Results appear only in papers. Inferred32212
More

No leaderboard or challenge for these two tasks. Results live in papers.

Checked the project page, the habitat-lab README and aihabitat.org/challenge, which redirects to the 2023 HomeRobot OVMM challenge.

Code licence
MIT89
More

habitat-lab LICENSE: MIT, 'Copyright (c) Meta Platforms, Inc. and its affiliates'. habitat-sim is MIT per the GitHub API.

Data licence
CC BY-NC 4.01823
More

Task episodes (hab3_episodes) and benchmark assets (hab3_bench_assets): CC BY-NC 4.0.

Asset licence
Non-commercial for the HSSD scenes and the avatars242521+1
More

Scenes (HSSD) CC BY-NC 4.0. Humanoid avatars non-commercial, with conflicting labels. Spot model by permission of Boston Dynamics. YCB objects CC BY 4.0.

HSSD scenes · CC BY-NC 4.0. The card asks users to acknowledge the licence.24

Humanoid avatars: conflict · Card metadata says CC BY-NC-SA 4.0; the card text says the 12 avatars and reaching poses are CC BY-NC 4.0, and the walking motion is under the SMPL Body Motion File License.25

Spot model · Licence 'other': Boston Dynamics, all rights reserved; written permission for redistribution with attribution.21

YCB objects · CC BY 4.026

Access
Open. It can be downloaded without an account.182728+1
More

Code on GitHub; episodes, avatars and scenes on Hugging Face, downloadable without an account through habitat-sim's download script.

The HSSD card shows a licence acknowledgement prompt; the Hub API reports the dataset as not gated (2026-10-10).

Commercial use
Not allowed Inferred241825
More

Scenes, episodes and avatars are licensed for non-commercial use only. Not legal advice.

Published at
ICLR 202429
More

ICLR 2024 (poster). arXiv has only v1.

Sources 29

  1. 1Habitat 3.0: A Co-Habitat for Humans, Avatars and Robots (arXiv abstract page)Paper · 19 Oct 2023 · checked 10 Oct 2026
  2. 2Habitat 3.0 paper, full text v1 (Sections 3 to 5, Appendices A, D, E, F, G, H)Paper · 19 Oct 2023 · checked 10 Oct 2026
  3. 3Habitat 3.0 project pageOfficial site · Oct 2023 · checked 10 Oct 2026
  4. 4Semantic Scholar API: citations of arXiv:2310.13724 with contexts (342 records)Index · 10 Oct 2026 · checked 10 Oct 2026
  5. 5Following the Human Thread in Social Navigation (SDA), Table 1Paper · 17 Apr 2024 · checked 10 Oct 2026
  6. 6Robots Ask the Way: Communication-Enabled Social Navigation (Habitat 3.0c)Paper · 1 Jul 2026 · checked 10 Oct 2026
  7. 7CST-WM: A Causally Structured World Model for Embodied Visual TrackingPaper · 5 Sep 2026 · checked 10 Oct 2026
  8. 8habitat-lab LICENSERepository · 2019 · checked 10 Oct 2026
  9. 9GitHub API: facebookresearch/habitat-sim (stars, licence)Index · 10 Oct 2026 · checked 10 Oct 2026
  10. 10Introducing Habitat 3.0: The next milestone on the path to socially intelligent robotsOfficial blog · 20 Oct 2023 · checked 10 Oct 2026
  11. 11habitat-baselines README: Social Navigation and Social Rearrangement sectionsRepository · 2026 · checked 10 Oct 2026
  12. 12habitat-lab README (maintenance notice)Repository · 7 May 2026 · checked 10 Oct 2026
  13. 13habitat-lab issue #1839: cannot evaluate the released social_nav_latest.pth checkpointRepository · 7 Mar 2024 · checked 10 Oct 2026
  14. 14habitat-lab issue #2105: object handles missing in provided episodesRepository · 8 Nov 2024 · checked 10 Oct 2026
  15. 15habitat-lab issue #1684: IndexError while evaluating baseline social navRepository · 14 Nov 2023 · checked 10 Oct 2026
  16. 16facebookresearch/habitat-lab releasesRepository · 7 May 2026 · checked 10 Oct 2026
  17. 17facebookresearch/habitat-sim releasesRepository · 12 Feb 2026 · checked 10 Oct 2026
  18. 18ai-habitat/hab3_episodes dataset cardDataset page · 28 Nov 2023 · checked 10 Oct 2026
  19. 19habitat-lab commit history (main)Repository · 7 May 2026 · checked 10 Oct 2026
  20. 20habitat-sim README (maintenance notice)Repository · 7 May 2026 · checked 10 Oct 2026
  21. 21ai-habitat/hab_spot_arm dataset card (Spot URDF licence note)Dataset page · 14 Feb 2025 · checked 10 Oct 2026
  22. 22AI Habitat challenge page (redirects to the 2023 HomeRobot OVMM challenge)Official site · 2023 · checked 10 Oct 2026
  23. 23ai-habitat/hab3_bench_assets dataset cardDataset page · 22 Feb 2024 · checked 10 Oct 2026
  24. 24hssd/hssd-hab dataset cardDataset page · 14 Feb 2025 · checked 10 Oct 2026
  25. 25ai-habitat/habitat_humanoids dataset cardDataset page · 18 Oct 2023 · checked 10 Oct 2026
  26. 26ai-habitat/ycb dataset cardDataset page · unknown · checked 10 Oct 2026
  27. 27Hugging Face Hub API record for ai-habitat/hab3_episodes (downloads)Index · 10 Oct 2026 · checked 10 Oct 2026
  28. 28Hugging Face Hub API record for hssd/hssd-hab (downloads, gated flag)Index · 10 Oct 2026 · checked 10 Oct 2026
  29. 29ICLR 2024 poster page: Habitat 3.0: A Co-Habitat for Humans, Avatars, and RobotsPaper · May 2024 · checked 10 Oct 2026
Where we searched for missing information

sim_to_real: Habitat 3.0 paper (full text, including Appendix D on the Spot robot and Appendix H), project page, Meta launch blog, SDA (2404.11327), Falcon (2409.13244), CommNav (2607.01044), Semantic Scholar citation contexts for all 342 citing papers filtered for real-robot or hardware mentions, web searches for real Spot deployment of Habitat 3.0 social navigation policies. None found.

leaderboard: Project page, habitat-lab and habitat-baselines READMEs, aihabitat.org/challenge (redirects to the 2023 HomeRobot OVMM challenge).

used_by: Semantic Scholar citation contexts (342 records) scanned for the task metrics (SPS, following rate, relative efficiency, zero-shot coordination); web searches for papers reporting these metrics.

trials (number of test episodes used for the paper's tables): Paper Sections 4.1 and 4.2 and Appendices A and E; habitat-baselines README; hab3_episodes card. The figure and table captions do not name the split.

issues (reproduction problems): habitat-lab GitHub issues searched for social_nav, social rearrangement and hab3 episodes. OpenReview reviews could not be read (the API returned HTTP 403).

Change history

  1. Created at full depth from primary sources, starting from the basic entry and research/raw/inventory/core-sim-b.json. Changes from the basic entry: leaderboard set to paper-only (was unknown with a 'papers only' display); added the human-in-the-loop fact, evaluation-episode conflict, licence conflict for the humanoid avatars, derived benchmarks, used_by, issues and readings. The basic entry's '1191±3 FPS' figure was not found in the text; we use the intro's 1190 FPS. Checks ran on 2026-10-10 and into early 2026-10-11 local time.
  2. Published as a full entry.