Overview
SharedAssembly is a novel shared-autonomy bilateral teleoperation framework designed to facilitate data collection for training tactile-aware robotic foundation models. This system embeds assembly-specific intelligence on both the leader and follower sides of the teleoperation setup. The framework addresses the operational barriers associated with conventional teleoperation in acquiring high-precision, tight-clearance assembly demonstrations.
Research Context
The development of tactile-aware robotic foundation models requires extensive demonstrations, particularly those involving high-precision and tight-clearance assembly tasks. The acquisition of such demonstrations is currently bottlenecked by the high operational demands and challenges inherent in conventional teleoperation systems. These systems often require significant operator expertise and can be inefficient for large-scale data collection. The need for scalable data collection methods in contact-rich manipulation motivates the development of frameworks like SharedAssembly.
Approach
The researchers proposed SharedAssembly as a shared-autonomy bilateral teleoperation framework. This framework incorporates assembly-specific intelligence on both the leader (human operator interface) and follower (robot manipulator) sides. The objective of this design is to streamline the teleoperation process and reduce the operational burden, thereby making data collection more efficient and accessible. The framework was evaluated through real-world user studies involving challenging sub-millimeter assembly tasks.
Findings
- SharedAssembly achieved a 97% assembly success rate in rigorous real-world user studies on challenging sub-millimeter tasks.
- The framework significantly boosted completion efficiency during these assembly tasks.
- The performance gains observed with SharedAssembly became more pronounced as the assembly clearance decreased.
- SharedAssembly effectively eliminated the expertise gap, allowing novice operators to achieve performance levels comparable to or exceeding those of expert operators using conventional systems.
- The framework offers an efficient, robust, and accessible solution for large-scale data harvesting in contact-rich manipulation by minimizing the skill barrier.
Why This Matters
This framework provides a solution for scalable data collection in contact-rich manipulation by reducing the skill barrier for operators. It facilitates the acquisition of indispensable high-precision, tight-clearance assembly demonstrations needed for training tactile-aware robotic foundation models. The ability for novice operators to achieve expert-level performance makes the data collection process more accessible and potentially widespread.