ASID: Active Exploration for System Identification in Robotic Manipulation

Authors: Marius Memmel, Andrew Wagenmaker, Chuning Zhu, Dieter Fox, Abhishek Gupta

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

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct our experimental evaluation in two scenarios. First, we conduct empirical experiments entirely in simulation (Todorov et al., 2012) to validate the behavior of the exploration, system identification, and downstream task modules of ASID. This involves treating a true simulation environment as the real-world environment and then a reconstruction of this simulation environment as the approximate constructed simulation. Policies learned in this approximate simulation can then be evaluated back in the original simulation environment to judge efficacy. Second, we apply this to two real-world manipulation tasks that are inherently dependent on accurate parameter identification, showing the ability of ASID to function in the real world, using real-world data.
Researcher Affiliation Academia Paul G. Allen School of Computer Science & Engineering University of Washington Seattle, WA 98195, USA {memmelma,ajwagen,zchuning,fox,abhgupta}@cs.washington.edu
Pseudocode No The paper describes its methodology in prose and figures, but does not provide pseudocode or a clearly labeled algorithm block.
Open Source Code Yes Project website at https://weirdlabuw.github.io/asid
Open Datasets No The paper conducts experiments in custom-built simulated environments (MuJoCo) and collects its own real-world data, but does not state that these datasets are publicly available or provide access information for them.
Dataset Splits No The paper describes parameter randomization ranges used during training and evaluation within the simulated environments, but does not specify explicit train/validation/test dataset splits (e.g., percentages or sample counts) for any fixed dataset.
Hardware Specification No The paper mentions robotic hardware used for real-world experiments (Franka Emika Panda robot, Intel Real Sense D455 cameras) but does not specify the hardware (e.g., GPU/CPU models, memory) used for training the models.
Software Dependencies No The paper names several software components and algorithms used (e.g., PPO, CEM, REPS, MuJoCo), but does not provide specific version numbers for these dependencies.
Experiment Setup No The paper describes parameter randomization ranges and choices of optimization algorithms (PPO, CEM, REPS) but does not provide specific hyperparameters like learning rates, batch sizes, or detailed training configurations for these algorithms.