Self Supervised Detection of Incorrect Human Demonstrations: A Path Toward Safe Imitation Learning by Robots in the Wild

A major appeal of learning from demonstrations or imitation learning (IL) in robotics is that it learns a policy directly from lay users. However, Lay users may inadvertently provide erroneous demonstrations that lead to learning of policies that are inaccurate and hence, unsafe for humans and/or robot. This paper makes two contributions in the endeavour of recognizing human errors in demonstrations and thereby helping to learn a safe IL policy. First, we created a dataset – Layman V1.0 – with 15 lay users who provided a total of 1200 demonstrations for three simulated tasks – Lift, Can and Square in the simulated Robosuite environment – and two real robot tasks with a Sawyer robot, using a custom designed Android app for tele-operation. Second, we propose a framework named Behavior Cloning for Error Detection (BED) to autonomously detect and discard erroneous demonstrations from a demonstration pool. Our method uses a Behavior Cloning method as self-supervised technique and assigns binary weight to each demonstration based on its inconsistencies with the rest of the demonstrations. We show the effectiveness of this framework in detecting incorrect demonstrations in the Layman V1.0 dataset. We further show that state-of-the-art (SOTA) policy learners learns a better policy when bad demonstrations, identified through the proposed framework, are removed from the training pool. Dataset and Codes are available in https://github.com/AssistiveRoboticsUNH/bed