Habitat 2.0: Training Home Assistants to Rearrange their Habitat (2106.14405v2)

Published 28 Jun 2021 in cs.LG and cs.RO

Abstract: We introduce Habitat 2.0 (H2.0), a simulation platform for training virtual robots in interactive 3D environments and complex physics-enabled scenarios. We make comprehensive contributions to all levels of the embodied AI stack - data, simulation, and benchmark tasks. Specifically, we present: (i) ReplicaCAD: an artist-authored, annotated, reconfigurable 3D dataset of apartments (matching real spaces) with articulated objects (e.g. cabinets and drawers that can open/close); (ii) H2.0: a high-performance physics-enabled 3D simulator with speeds exceeding 25,000 simulation steps per second (850x real-time) on an 8-GPU node, representing 100x speed-ups over prior work; and, (iii) Home Assistant Benchmark (HAB): a suite of common tasks for assistive robots (tidy the house, prepare groceries, set the table) that test a range of mobile manipulation capabilities. These large-scale engineering contributions allow us to systematically compare deep reinforcement learning (RL) at scale and classical sense-plan-act (SPA) pipelines in long-horizon structured tasks, with an emphasis on generalization to new objects, receptacles, and layouts. We find that (1) flat RL policies struggle on HAB compared to hierarchical ones; (2) a hierarchy with independent skills suffers from 'hand-off problems', and (3) SPA pipelines are more brittle than RL policies.

Authors (21)

Andrew Szot (15 papers)
Alex Clegg (3 papers)
Eric Undersander (11 papers)
Erik Wijmans (25 papers)
Yili Zhao (4 papers)
John Turner (7 papers)
Noah Maestre (2 papers)
Mustafa Mukadam (43 papers)
Devendra Chaplot (4 papers)
Oleksandr Maksymets (17 papers)
Aaron Gokaslan (33 papers)
Sameer Dharur (6 papers)
Franziska Meier (46 papers)
Wojciech Galuba (9 papers)
Angel Chang (5 papers)
Zsolt Kira (110 papers)
Vladlen Koltun (114 papers)
Jitendra Malik (211 papers)
Manolis Savva (64 papers)
Dhruv Batra (160 papers)

Citations (434)

View on Semantic Scholar

Summary

An Expert Overview of "Habitat 2.0: Training Home Assistants to Rearrange their Habitat"

The paper "Habitat 2.0: Training Home Assistants to Rearrange their Habitat" presents a pivotal advancement in simulation platforms for embodied AI research, focusing on virtual robots in dynamic 3D environments. This work encompasses contributions across data, simulation, and benchmarking, crucial for developing and testing AI systems in controlled yet comprehensive settings.

Key Contributions

ReplicaCAD Dataset: This dataset represents a meticulously designed collection of 3D apartment models, complete with movable objects like cabinets and drawers. Comprising 111 unique layouts and 92 dynamic objects, this dataset facilitates studies on generalization in varied home environments.
Habitat 2.0 Simulator: Habitat 2.0 is a high-performance simulation environment capable of executing 25,000 simulation steps per second, offering a significant speed advantage over predecessors. This speed enables efficient reinforcement learning at scale and reduces experimental cycles drastically, thus fostering feasibility for extensive, long-horizon tasks.
Home Assistant Benchmark (HAB): A suite of tasks designed to evaluate mobile manipulation capabilities in assistive robots. This setup focuses on real-world applications like arranging groceries or setting a table, with challenges directed at both reinforcement learning and classical approaches.

Findings

The paper's experiments reveal several insights into reinforcement learning and classical robotics approaches:

Flat vs. Hierarchical RL Policies: Hierarchical RL policies outperform flat ones, particularly in intricate tasks requiring skill chaining. The paper highlights challenges in crafting reward functions that facilitate seamless skill transitions.
SPA (Sense-Plan-Act) Pipeline Robustness: Classical SPA methods exhibit brittleness in perceiving complex, cluttered environments. The limitations in situational mapping and planning from partial observations make them less robust compared to RL policies.
Generalization: The experiments underscore challenges in generalizing RL policies to unseen objects and environments, pointing to the need for diverse training datasets and scenarios.

Implications

The implications of this work span both theoretical and practical dimensions:

Theoretical: The work enhances understanding of embodied AI, particularly in how reinforcement learning can be structured and maximized for tasks involving dynamic environments and long time horizons.
Practical: The flexible, scalable simulation environment and dataset provide a powerful tool for real-world robotics applications, drastically reducing development time and enabling reproducible, comprehensive testing.

Future Directions

The paper lays groundwork for future exploration in several areas:

Expanding Dataset Diversity: Increasing the cultural and structural diversity of environments can enhance generalization of AI models across global contexts.
Integration of Advanced Functions: Incorporating non-rigid object dynamics and other complex interactions remains a promising yet unexplored frontier within Habitat 2.0.
Holistic Optimization: There is potential for further optimizing the interaction between simulation, rendering, and RL processes to enhance throughput and fidelity.

In summary, "Habitat 2.0" constitutes a significant stride forward in simulation for embodied AI, providing a robust framework for both research and practical applications in training home assistants. The insights gained from this work will likely propel further advancements in AI and robotics, with extensive possibilities for future exploration.

PDF Markdown