The Next Generation of Robotic Hands
ETH Zurich’s Soft Robotics Lab introduces the Orca hand, a tendon-driven, sensorized robotic hand designed to bring humanlike dexterity within reach of everyday research labs. Built from pop-in pin joints and assembled in under eight hours for about $2,000, Orca offers durability, easy repair, and impressive accuracy rivaling far bulkier direct-drive systems. Reliability tests show continuous gripping for hours without overheating or slack issues, thanks to a low-friction tendon design and a smart retensioning mechanism.
The team showcases Orca’s autonomy with long-horizon pick-and-place loops and robust reinforcement-learning policies that enable tasks like tennis-ball reorientation. In short: a fast-to-build, resilient, high-performance robotic hand that lowers the barrier to real dexterous manipulation research.
Transcript
00:00:01 Generalpurpose robots must possess humanlike dexterity and agility to match the versatility of humans. A humanlike form factor further enables leveraging the wealth of data available from human hand interactions. However, the bottleneck in dextrous manipulation lies not only in software but arguably even more in hardware. Robotic hands matching human capabilities are often
00:00:24 prohibitively expensive, bulky, or require enterprise level maintenance, making them inaccessible for broader research and applications. What if the research community could get started with reliable dextrous hands within a day? We present the Orca hand, a tendon-driven robotic hand with fully integrated tactile sensors that can be assembled in under 8 hours and with a
00:00:47 material cost basis of only $2,000. The popable pin joint design does not only allow for a quick assembly, but it also prevents joints from breaking when excessive stress is applied. The reusability of the joint after failure saves valuable time and resources that would otherwise have been wasted on rebuilding the necessary parts. After successful assembly of the Orca hand, an
00:01:11 autoc calibration step will be performed by testing the maximal allowed joint angles. The hand is afterwards ready for precise movement control. We benchmarked our systems tendon-driven dynamics against the dynamics of a direct driven robot hand and were able to achieve similar accuracy to the leap hand while being far less bulky in our design.
00:01:34 A typical downside to tendonbased systems is their reliability over time. In our reliability tests, we could let the Orca hand perform continuous gripping for over 2.5 hours, at which point we terminated the experiment without signs of overheating or performance decrease. This great durability is thanks to a design of our tendon system with minimal friction as
00:01:56 well as a manual retensioning ratchet spool mechanism. To further the autonomy of the Orca hand beyond teley operation, we designed a continuous pick and place evaluation task. The experiment requires the hand, which is now equipped with a camera, to pick up a cube and drop it onto a sloped surface. From there, it can roll into a new random position, and the experiment
00:02:20 can repeat itself. To achieve this, the experiment was separated into a training phase in which a teley operator continuously performed this task and a test phase in which the robot tried to imitate what it had seen. We were able to deploy the robot on this task for over seven hours without the need for human intervention or failure due to tendon slack.
00:02:43 Lastly, we trained the hand with reinforcement learning, which can be useful for tasks that are hard to perform yourself. After a 1-hour training epic, we could deploy a robust policy that allowed us to reorient a tennis ball with the Orca hand.