[Paper Review] Robot Learning in Homes: Improving Generalization and Reducing Dataset Bias

Abhinav Gupta, Adithyavairavan Murali+2 more|arXiv (Cornell University)|Jul 18, 2018

Robot Manipulation and Learning61 citations

TL;DR

This paper collects a large 28k grasp dataset in six real homes using a low-cost robot, introduces a noise-robust learning architecture to model label noise, and shows substantial generalization gains over lab-trained baselines (43.7% improvement) and DexNet baselines, plus a 10% benefit from noise modeling.

ABSTRACT

Data-driven approaches to solving robotic tasks have gained a lot of traction in recent years. However, most existing policies are trained on large-scale datasets collected in curated lab settings. If we aim to deploy these models in unstructured visual environments like people's homes, they will be unable to cope with the mismatch in data distribution. In such light, we present the first systematic effort in collecting a large dataset for robotic grasping in homes. First, to scale and parallelize data collection, we built a low cost mobile manipulator assembled for under 3K USD. Second, data collected using low cost robots suffer from noisy labels due to imperfect execution and calibration errors. To handle this, we develop a framework which factors out the noise as a latent variable. Our model is trained on 28K grasps collected in several houses under an array of different environmental conditions. We evaluate our models by physically executing grasps on a collection of novel objects in multiple unseen homes. The models trained with our home dataset showed a marked improvement of 43.7% over a baseline model trained with data collected in lab. Our architecture which explicitly models the latent noise in the dataset also performed 10% better than one that did not factor out the noise. We hope this effort inspires the robotics community to look outside the lab and embrace learning based approaches to handle inaccurate cheap robots.

Motivation & Objective

• Demonstrate that collecting robotic grasp data in diverse home environments reduces dataset bias and improves generalization.
• Show that low-cost robots introduce label noise and that this noise can be explicitly modeled to improve learning.
• Provide a scalable data collection and learning framework that generalizes to unseen homes and objects.

Proposed method

• Assemble a mobile manipulator under $3k to collect 28k grasp examples in six homes.
• Model noisy labels as a latent variable and train a Robust-Grasp architecture combining a Grasp Prediction Network (GPN) and a Noise Modelling Network (NMN) with a marginalization layer.
• Patch-based grasp formulation: sample nine image patches around a candidate grasp and predict success probabilities for each patch and angle.
• Model noise via NMN using global scene information, robot ID, and grasp pixel location; train NMN and GPN jointly without explicit EM labels.
• Two-stage training: first train GPN with noisy patches, then jointly train NMN and GPN with end-to-end optimization.
• Utilize pre-trained ResNet-18 features and train with Adam; evaluate on held-out data and real hardware (Real-LCA and Real-Sawyer).

Experimental results

Research questions

• RQ1 Do home-collected grasp datasets improve generalization to unseen home environments compared to lab-collected data?
• RQ2 Can a latent-noise modeling approach improve learning from noisy data collected by low-cost robots?
• RQ3 How does Robust-Grasp perform on novel objects/environments compared to DexNet baselines and Patch-Grasp?
• RQ4 What is the impact of training with home data on lab-hosted and home deployments?

Key findings

• Robust-Grasp trained on Home-LCA achieves higher held-out accuracy (73.0%) than Patch-Grasp (69.9%) when tested on Home-LCA, and outperforms lab-based training baselines.
• Training on Home-LCA yields a 43.7% improvements over a model trained on Lab-Baxter data when evaluated on Real-LCA hardware.
• Robust-Grasp trained on Home-LCA is about 33% better than DexNet on Real-LCA scenes, highlighting robustness in non-ideal home sensing.
• Factoring out latent noise in data yields roughly a 10% performance improvement over a model that does not model noise, as shown in Real-Sawyer tests (77.5% vs 56.25%).
• Dataset: about 28,000 grasps collected across six homes; testing on three unseen homes demonstrates better generalization to real home environments.