Perceptyne Thesis
1 Introduction
Automation has been extremely successful for mass production of objects with little to no variation. However, automation has been difficult in industries where environments/objects are complex/unstructured. For example, even today, tasks like product assembly (mobile phones, power banks, automotive headlights, TVs etc.), product packaging, food preparation & packing, garment preparation etc. are all mostly done manually. One illustrative fact that shows the magnitude of the dependence on manual labor is that, even at the scale of Apple, every single iPhone is still manually assembled, somewhere in labor centric economies. Often, these tasks are repetitive, dreary, and prone to errors. Technically, all the above tasks fall under the category of what we call dexterous tasks, and are quite hard to automate economically, while retaining speed and flexibility. Automation solutions today are often custom built, need lots of scenario specific programming, and are quite expensive. And almost always, they cannot accommodate changes to product / workflow.
While robotic arm-based automation is a step in the right direction, current day robots still lack many key capabilities to be as dexterous as human operators on a typical assembly line. The result is: even today, most repetitive industrial tasks still depend on manual effort.
Our vision is to build generic, low cost, dexterous industrial robots that can help automate these tasks, scalable. In this approach paper, we outline the limitations of current day robot platforms, and what it takes to achieve highly dexterous, low-cost, contextually intelligent robots for automating these repetitive tasks.
2 What are dexterous tasks?
Dexterous tasks are tasks that need hand-eye coordination along with a fine sense of touch, grasp and object manipulation, in addition to a comprehensive model of how the physical world works. Some videos of these dexterous tasks are shown here, and here[2]. Following activities are some examples of dexterous tasks:
· Putting together a Wi-Fi router, from its components
· Packing a product that needs tube shipped in a carton box
· Preparing a Subway sandwich, or a McDonalds burger
· Putting together a meal tray for consumption on an airline
Fig.1: Example dexterous tasks in product assembly and packing [3]
3 Limitations of current day robots
Robots today are severely limited when it comes to dexterous tasks. One cannot imagine a dual-arm robot capable of skillfully chopping vegetables or carefully putting together a mobile phone or even packing a cardboard carton for a shipment. A few key limitations of today’s robots in the context of these tasks are out lined in the sections below:
3.1 Lack dexterous end-effectors (grippers)
There are mainly4 types of robotic grippers today: Vacuum grippers, 2 finger grippers, 3 finger claspers, flexible grippers. All these grippers usually somehow hold an object that is conducive enough. They struggle in getting a good, preferred orientation grasp of the object. Often, it is the gripper’s limitation (rather than the application) that decides the actual orientation of grasping the object. A video highlighting the level of dexterity needed for product assembly can be seen here, and it is evident how inadequate typical grippers are for dexterous tasks.
3.2 Expensive
A typical articulate robot with 5 kg payload costs upwards of 25k USD. If we need a dual arm configuration, along with accessories, the cost becomes ~70k USD, which including software would land up at ~100k USD. To put this in perspective, typical salary of a line operator in a labor centric economy like China would be about 5-10k USD per annum. This is one of the fundamental drivers for outsourcing manufacturing to China. For robot deployments to make extensive economic sense (both in developed & developing economies), they must drop down from the current 100k USD figure to ~20k USD figure. That enhances the incentive to automate the tasks, freeing human potential for better tasks.
3.3 Custom built jigs & rigs | Inflexible
Consider this video showcasing a special purpose machine that assembles a torch light [4]. While it is based on a robotic arm, the entire setup is quite inflexible in its functionality. It is a custom rig that can assemble a torch light, and nothing other than a torch light. If the size of the torch light varies a bit, a new jig must be made. If the external design of the torch light changes, significant changes may be needed to the jigs. On the other hand, humans have the flexibility to assemble just about any product with the same pair of hands & eyes. This is the kind of flexibility robots today do not possess.
3.4 The problem of instruction
Robots today excel in structured position-controlled environments. The reason for this is that most robots are primarily programmed with sequences of kinematic maneuvers, i.e., instructions that look like:
move_to(x_1,y_1,z_1 )→close_gripper()→ move_to(x_2,y_2,z_2 )→release_gripper() While this approach is simple and robust, it cannot scale to most unstructured environments. For example, it would be quite difficult to write a sequence of kinematic maneuvers, to codify the task of chopping vegetables. Tele-operatively trained robots, with a sense of hierarchical task planning and outcome based sub-task execution, work the best for these scenarios.
3.5 Slow to train themselves.
Even robots that do use AI, need far more shots than humans to learn & replicate simple repetitive tasks with relatively narrow scope. A prime reason for this is that robots today rely only on machine vision and do not exploit other sensory modalities like tactile sensing, force feedback and depth sensing. Usage of multi-modal sensing significantly speeds up the task of training robots to intelligently replicate dexterous actions.
4 Our approach
Our core thesis:
a. Purpose-built automation solutions have reached their saturation. The future is of general-purpose, AI-driven, dexterous robots.
b. These robots are possible only by tight integration of high-DoF manipulators, multi-modal sensing & appropriate AI paradigms.
In the above context, we have highlighted the key technological elements that form a defining part of our platform and are essential for dexterity. They are:
· Robotic hands with:
o High-DoF (Degree of Freedom), Agility& Anthropomorphism*
o Tactile sensing
o Hard-Soft structure, hermetically sealed
· Robotic arms that have:
o Dual-arm configuration
o Force feedback
o Visual servo capability
o Low-cost gearboxes for robot joints
· Tele-operative, AI assisted training methodology.