Habitat 2.0: Training Home Assistants to Rearrange their Habitat
Authors: Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimír Vondruš, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, Dhruv Batra
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct a systematic study of two distinct techniques monolithic sensors-to-actions policies trained with 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. |
| Researcher Affiliation | Collaboration | 1Facebook AI Research, 2Georgia Tech, 3Intel Research, 4Simon Fraser University 5UC Berkeley |
| Pseudocode | No | The paper describes methods and algorithms in prose but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code and setup instructions can be found at https://github.com/facebookresearch/habitat-lab. |
| Open Datasets | Yes | Our starting point was Replica [2], a dataset of highly photo-realistic 3D reconstructions at room and building scale. ... These inserted objects can come from Replica CAD or the YCB dataset [59]. |
| Dataset Splits | No | The paper describes training on certain objects, receptacles, and layouts, and evaluating on 'unseen' counterparts for systematic generalization, but it does not specify traditional dataset splits (e.g., 80/10/10) for training, validation, and test sets. Instead, it refers to 'episodes' for evaluation. |
| Hardware Specification | Yes | Benchmarking was done on machines with dual Intel Xeon Gold 6226R CPUs 32 cores/64 threads (32C/64T) total and 8 NVIDIA Ge Force 2080 Ti GPUs. |
| Software Dependencies | Yes | We used python-3.8 and gcc-9.3 for compiling H2.0. |
| Experiment Setup | Yes | Each simulation step consists of 1 rendering pass and 4 physics-steps, each simulating 1/120 sec for a total of 1/30 sec. New joint position goals are set every 1/30 sec and a joint controller computes the joint torques to achieve the joint goals for the current joint state every 1/120 sec. |