End-to-End Egospheric Spatial Memory

Authors: Daniel James Lenton, Stephen James, Ronald Clark, Andrew Davison

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through our broad set of experiments, we show that ESM provides a general computation graph for embodied spatial reasoning, and the module forms a bridge between real-time mapping systems and differentiable memory architectures.
Researcher Affiliation Collaboration Daniel Lenton 1, Stephen James 1, Ronald Clark 2, Andrew J. Davison 1 1 Dyson Robotics Lab, 2 Department of Computing, Imperial College London
Pseudocode Yes Algorithm 1: ESM Step
Open Source Code Yes Implementation at: https://github.com/ivy-dl/memory.
Open Datasets Yes For the object segmentation experiment, we use downsampled 60 80 and 120 160 images from the Scan Net dataset, which we first RGB-Depth align.
Dataset Splits Yes The losses for each network evaluated on the training set and validation set during the course of training are presented in Fig 12.
Hardware Specification Yes running on Nvidia RTX 2080 GPU
Software Dependencies Yes The implementation of our module is therefore jointly compatible with Tensor Flow 2.0, Py Torch, MXNet, Jax and Numpy.
Experiment Setup Yes All networks use a batch size of 16, an unroll size of 16, and are trained for 250k steps using an ADAM optimizer with 1e 4 learning rate.