Instrumentation for Better Demonstrations: A Case Study

TL;DR

The paper addresses data efficiency in learning from demonstrations by showing how instrumenting a manipulation task with sensors can automate demonstrations and provide privileged information. It presents a case study where a sensorized squeeze bottle, a PI controller, and transformer-based policies are used to learn a constant-flow liquid dispensing task, compared against human demonstrations. The key finding is that policies trained on automated PI demonstrations outperform those trained on human demonstrations in about 78% of cases, illustrating both scalable data collection and improved policy quality. This work highlights instrumentation as a promising path toward scalable, generalist robotic agents and motivates further studies on generalization and broader instrumentation strategies.

Abstract

Learning from demonstrations is a powerful paradigm for robot manipulation, but its effectiveness hinges on both the quantity and quality of the collected data. In this work, we present a case study of how instrumentation, i.e. integration of sensors, can improve the quality of demonstrations and automate data collection. We instrument a squeeze bottle with a pressure sensor to learn a liquid dispensing task, enabling automated data collection via a PI controller. Transformer-based policies trained on automated demonstrations outperform those trained on human data in 78% of cases. Our findings indicate that instrumentation not only facilitates scalable data collection but also leads to better-performing policies, highlighting its potential in the pursuit of generalist robotic agents.

Figures (4)

• Figure 1: Experimental setup for learning to squeeze a constant flow from a bottle based on demonstrations, collected either by a human or by an autonomous controller relying on instrumentation data.
• Figure 2: Control flow diagram for data collection.
• Figure 3: A constant flow requires non-linear squeezing.
• Figure 4: Comparing agent scores.