Entity-Centric Reinforcement Learning for Object Manipulation from Pixels

Authors: Dan Haramati, Tal Daniel, Aviv Tamar

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method on several simulated tabletop robotic object manipulation environments implemented with Isaac Gym (Makoviychuk et al., 2021). The environment includes a robotic arm set in front of a table with a varying number of cubes in different colors.
Researcher Affiliation Academia Dan Haramati, Tal Daniel, Aviv Tamar Department of Electrical and Computer Engineering, Technion Israel Institute of Technology {haramati, taldanielm}@campus.technion.ac.il; avivt@technion.ac.il
Pseudocode No The paper describes its methods in detail and includes an outline of the architecture (Figure 2), but it does not contain explicit pseudocode blocks or algorithms labeled as such.
Open Source Code Yes Our code is publicly available on https://github.com/Dan Hrmti/ECRL.
Open Datasets No We collect 600,000 images from 2 viewpoints by interacting with the environment using a random policy for 300,000 timesteps.
Dataset Splits No The paper describes collecting data via interaction with a simulated environment and training models, but it does not explicitly provide details on how this collected data is split into training, validation, and test sets with specific percentages or counts.
Hardware Specification No The paper states that environments are implemented with Isaac Gym and makes general references to GPU-based simulation, but it does not specify concrete hardware details such as GPU models, CPU types, or memory used for experiments.
Software Dependencies No We implement our RL algorithm with code adapted from stable-baselines3 (Raffin et al., 2021). We train a DLPv2 (Daniel & Tamar, 2023) using the publicly available code base. The paper mentions frameworks and libraries with their publication years but does not provide specific version numbers required for exact replication (e.g., `stable-baselines3==x.y.z`).
Experiment Setup Yes Further details and hyper-parameters can be found in Appendix D. Our code is publicly available on https://github.com/Dan Hrmti/ECRL. The paper includes detailed tables such as “Table 7: General hyper-parameters used for RL training.” and “Table 9: Hyper-parameters for the EIT architecture.” listing specific values for learning rates, batch sizes, attention dimensions, etc.