Ask4Help: Learning to Leverage an Expert for Embodied Tasks
Authors: Kunal Pratap Singh, Luca Weihs, Alvaro Herrasti, Jonghyun Choi, Aniruddha Kembhavi, Roozbeh Mottaghi
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate ASK4HELP on two different tasks object goal navigation and room rearrangement and see substantial improvements in performance using minimal help. On object navigation, an agent that achieves a 52% success rate is raised to 86% with 13% help and for rearrangement, the state-of-the-art model with a 7% success rate is dramatically improved to 90.4% using 39% help. Human trials with ASK4HELP demonstrate the efficacy of our approach in practical scenarios. |
| Researcher Affiliation | Collaboration | Kunal Pratap Singh PRIOR, Allen Institute for AI Luca Weihs PRIOR, Allen Institute for AI Alvaro Herrasti PRIOR, Allen Institute for AI Jonghyun Choi Yonsei University Aniruddha Kembhavi PRIOR, Allen Institute for AI Roozbeh Mottaghi PRIOR, Allen Institute for AI |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | No | The paper states 'We use the publicly available codebase5 provided by [25]' and provides a link to it, but this refers to the codebase of a *prior* model (Embodied CLIP) that they used, not the open-source code for their own ASK4HELP methodology. |
| Open Datasets | Yes | We train the Robo THOR [14] Object Nav and i THOR 1-phase Room R [52] models proposed in Emb CLIP [25]... The open-source Robo THOR [14] environment supports this task... |
| Dataset Splits | Yes | We train the Robo THOR [14] Object Nav and i THOR 1-phase Room R [52] models proposed in Emb CLIP [25], currently the published So TA models for these two tasks, on 75% of the training scenes for their respective tasks (45 scenes for Object Nav and 60 for Room R). ... We use the remaining 25% training scenes to train the the ASK4HELP policy. We evaluate our models on the unseen validation scenes that the agent has not seen before in training. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, memory) used to run the experiments. |
| Software Dependencies | No | The paper mentions using the 'Allen Act [53] framework' but does not specify any version numbers for this framework or other key software dependencies like programming languages or libraries. |
| Experiment Setup | Yes | We train the ASK4HELP policy using DD-PPO [54] for 15 million steps. ... For Object Nav, we set rfail = 10, rinit_ask = 1, and rstep_ask = 0.01. ... we vary, rfail from -1 to -30 to cover a broad range of user preferences and generate rewards configuration namely R 1 to R 30. |