Zero-Shot Visual Imitation

Authors: Deepak Pathak, Parsa Mahmoudieh, Guanghao Luo, Pulkit Agrawal, Dian Chen, Yide Shentu, Evan Shelhamer, Jitendra Malik, Alexei A. Efros, Trevor Darrell

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our zero-shot imitator in two real-world settings: complex rope manipulation with a Baxter robot and navigation in previously unseen office environments with a Turtle Bot. Through further experiments in Viz Doom simulation, we provide evidence that better mechanisms for exploration lead to learning a more capable policy which in turn improves end task performance. Videos, models, and more details are available at https://pathak22.github.io/zeroshot-imitation/.
Researcher Affiliation Academia UC Berkeley {pathak,parsa.m,michaelluo,pulkitag,dianchen,fredshentu,shelhamer,malik,efros,trevor}@cs.berkeley.edu
Pseudocode No No pseudocode or algorithm blocks were found.
Open Source Code Yes Videos, models, and more details are available at https://pathak22.github.io/zeroshot-imitation/.
Open Datasets No No specific link, DOI, repository name, or formal citation with authors/year was provided for publicly available or open datasets. The paper mentions using data from Nair et al. (2017) and collecting their own data but no clear access information.
Dataset Splits No No specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) was found.
Hardware Specification No The paper mentions a "Baxter robot" and a "Turtle Bot2 robot comprising of a wheeled Kobuki base and an Orbbec Astra camera" and a "powerful on-board laptop". However, no specific GPU/CPU models, processor types, or memory details are provided for the compute used for training or inference of the models.
Software Dependencies No The paper mentions "Adam (Kingma & Ba, 2015)" as an optimizer, and "Alex Net" and "Res Net-50 (He et al., 2016)" as architectures, but no specific software library names with version numbers (e.g., PyTorch 1.9) were found for reproducibility.
Experiment Setup Yes The loss weight of the forward model is 0.1, and the objective is minimized using Adam (Kingma & Ba, 2015) with learning rate of 5e 4. All models were trained with batch size 64, Adam Solver with 1e-4 learning rate, and landmark slices uniformly sampled between 5 to 15 action steps for each batch. The observations are 42x42 resolution, grayscale images with only one-time channel both for goal and current state.