Robotics scientists have been attempting to achieve authentic mastery in robotic hands for a while, yet this objective has proven challenging. Despite the progress in robotics, there are still limitations when achieving advanced mastery in robot hands.
While robots equipped with grippers and suction cups have proven effective in handling basic tasks like picking up and moving objects, they face significant challenges when it comes to performing more intricate actions such as assembling, inserting, reorienting, and packaging. Human operators have traditionally carried out these tasks exclusively due to the complexity and fine motor skills involved.
The challenges in achieving robotic mastery stem from the intricacies of sensing and manipulating objects in real-world scenarios. Objects can have unpredictable shapes, textures, and material properties, making it difficult for robots to perceive and interact with them accurately. Additionally, the variability in environmental conditions further complicates the task. Objects may need to be picked up from disorganised environments or inserted into tight spaces, requiring precise control and adaptability. Replicating this level of mastery in robot hands has proven to be a tricky and elusive goal for researchers.
However, recent progress in sensing technology and machine learning algorithms for analysing gathered data has sparked a rapid transformation in the field of robotic manipulation.
Recently, researchers at Columbia Engineering, supported by the U.S. National Science Foundation, have successfully showcased a remarkably agile robot hand. This advanced robotic hand integrates a sophisticated tactile sense with motor learning algorithms, enabling the robot to acquire new physical skills through repetitive practice.
To exhibit the capabilities of their robotic hand, the researchers selected a challenging manipulation task involving a non-uniformly shaped object. The task involved rotating the object extensively while ensuring a stable and secure grip. The hand must continually reposition certain fingers while keeping the rest of the fingers in place to maintain the object’s stability. Remarkably, the hand successfully completed the task without relying on visual feedback, relying solely on touch sensing.
The absence of visual dependency in object manipulation allows a hand the ability to operate in challenging lighting conditions that would confuse algorithms based on visual perception. It can even perform tasks in complete darkness, further highlighting its versatility.
Mechanical engineer, Matei Ciocarlie, expressed that although the demonstration was conducted on a proof-of-concept task to showcase the hand’s capabilities, they firmly believe that such a high level of mastery will introduce entirely new applications for robotic manipulation in practical settings. Ciocarlie mentioned that immediate uses could include:
- Logistics and material handling.
- Addressing supply chain issues that have impacted the economy.
- Advanced manufacturing and assembly in factories.
Under the leadership of Gagan Khandate, the researchers designed and constructed a robot hand featuring five fingers and 15 independently controlled joints. The touch-sensing technology developed by the team was integrated into each finger. The subsequent step involved assessing the tactile hand’s aptitude for complex manipulation tasks. They employed a novel motor learning technique called deep reinforcement learning, and their newly devised algorithms facilitate the compelling exploration of various motor strategies.
Bruce Kramer, a program director in the Directorate for Engineering at the National Science Foundation (NSF), explained that the project, which received funding from the NSF’s Future Manufacturing Program, aims to explore novel avenues for leveraging machine intelligence in enabling robots to perform manufacturing tasks while adjusting to dynamic conditions autonomously.
The objective is to develop autonomous robots to learn from past experiences and proactively seek human assistance when their confidence level diminishes.
This approach opens new possibilities in the realm of manufacturing, where robots can operate independently and adapt to changing circumstances with the support of machine intelligence.