TLDR: This paper introduces a 4-channel bilateral control system for low-cost, force-sensorless manipulators, leveraging an accurate dynamics model to enable fast teleoperation with force feedback. It demonstrates superior position and force tracking compared to existing methods. Furthermore, it shows that incorporating force information from data collected via this system significantly improves the success rate of imitation learning tasks, making high-fidelity robot control and data collection accessible on affordable hardware.
Researchers have developed a groundbreaking method for controlling low-cost robotic manipulators, enabling them to perform complex tasks with high precision and force feedback, even without expensive force sensors. This advancement is particularly significant for collecting demonstration data for imitation learning, a field where robots learn by observing human actions.
Traditionally, low-cost robots often rely on “unilateral control,” which only sends position commands to the robot. While simple to implement, this approach struggles with tasks requiring delicate contact or high speed because it lacks force feedback. Imagine trying to pick up a fragile object or turn a stubborn nut without feeling how much force you’re applying – it’s incredibly difficult for a robot.
The new method introduces a “4-channel bilateral control” system. This system allows for two-way information exchange between a human operator’s “leader” robot and the “follower” robot. Both robots transmit their positions and forces to each other, enabling parallel position and force control. The key innovation lies in using an accurately identified dynamics model of the manipulator. This model helps compensate for nonlinear forces, estimate velocity and external forces, and adjust control gains based on the robot’s changing inertia. This means the robot can “understand” and react to forces, even without dedicated force sensors.
Experiments demonstrated the superior performance of this new system. When compared to unilateral control, symmetric position control, and force feedback control, the proposed 4-channel system showed significantly better position and force tracking during fast movements. For instance, in a rapid swinging motion, the new method drastically reduced position errors. The researchers also conducted “ablation studies” to show the importance of each component of their system, confirming that accurate inertia modeling, compensation for centrifugal and Coriolis forces, and a precise velocity observer are crucial for high performance.
Beyond teleoperation, the research also explored the impact on imitation learning. Data collected using this advanced 4-channel bilateral control was used to train a neural network model called Action Chunking with Transformer (ACT). The team investigated how incorporating force information into both the input (what the robot “sees” and “feels”) and output (what the robot “does”) of the learned policies affected performance.
The results were compelling. In tasks like dual-arm pick-and-place, nut turning, and cucumber peeling, including force information in the input significantly improved the success rate. For example, in the pick-and-place task, robots struggled with small objects without force input but achieved 100% success when force information was included. When force information was used for both input and output, stability and success rates further improved, especially for complex tasks like nut turning and cucumber peeling.
Also Read:
- Navigating Data Diversity: New Insights for Scalable Robotic Manipulation
- EC-Flow: Enabling Robots to Learn Complex Tasks from Unlabeled Videos
This research highlights a practical and effective way to achieve high-fidelity teleoperation and collect valuable, force-rich data using affordable hardware. This could accelerate the development of more capable and versatile robots for a wide range of applications. You can find the full research paper here.
