Welcome to the Bulletin by Remix Robotics, where we share a summary of the week's need-to-know robotics and automation news.

In today's email -

• Creepy spider robots

• Hearts for robot softies

• Self-driving is very expensive

• Should robots be flexible or specific

Snippets

Why why why why - Researchers at Rice University have found a way to use spiders as robot end effectors. In a new approach they term “necrobotics”, hydraulic pressure is used to actuate the legs of a dead spider, allowing the device to pick up objects. Despite claims by the lead academic that “this area of soft robotics is a lot of fun”, our personal opinion is that this is classic “Just because you can doesn’t mean you should”.

Good BEHAVIOR - A new paper called the “Benchmark for Everyday Household Activities in Virtual, Interactive and ecOlogical enviRonments” (with the questionable acronym ‘BEHAVIOR’) provides a virtual method for training and testing robots on 100 household tasks. It tests the ability of agents to perceive the environment, plan, and execute complex activities that involve multiple objects, and rooms. If it means we get a robot assistant sooner - we're happy. Find the benchmark for free here.

Soft-Hearted Robots - Soft robots use hydraulic pressure to actuate their limbs. However, this requires pumps, which aren’t flexible; this limits how ‘soft’ the whole system can be. Now, work by Cornell and the Army Research Lab has developed a small, in-line pump which acts as a heart for the robots. A magnet inside the soft tubing of the robot is surrounded by a magnetorheological fluid, which stiffens in the magnetic field into a solid plug. External solenoid coils can then be used to move the magnet, creating pressure. This pump performs much better than previous soft-robotics versions.

Red-Lining But Barely Moving - General Motors continues to post huge losses through its autonomous taxi service, Cruise: $5 million per day over the last quarter. Their CEO has a nice spin on it though - "When you’ve got the opportunity to go after a trillion-dollar market, you don’t casually wade into that," and "Aggressively pursuing the market is a competitive advantage.”.

Generally Useless - Robot costs are falling, but adoption is slower than expected — why? This article argues that it is due to robot designers creating general-purpose systems. Instead, we should start with the application (e.g welding or palletising) and build a specific solution to that challenge. Customers want holes, not drill bits. This is an interesting and counterintuitive take but truth is that many “special purpose machines” already exist. They’re generally expensive and tailored to stable high-volume processes. That said the future of robotics is preconfigured systems - whether it's due to flexible design or a special purpose design.

We preferred White Castle anyways - The CEO of Mcdonald's is not keen on kitchen robots - “It’s great for garnering headlines, it’s not practical in the vast majority of restaurants”. They’re much more interested in chatbot drive-throughs and in-app experiences.  On the other hand -White Castle has ordered 100 robot fry chefs.

The Big Idea

Building smart robots

It's no secret that industrial robots can be pretty ‘dumb’. Although flexible by design, when robots are installed in factories, they’re integrated to meet the specific needs of a specific factory. Robots are often hand programmed to follow predetermined trajectories and use a lot of custom tooling and fixtures. As such, the environment must be deterministic and with everything well defined in advance. Change in workpiece or process must be preprogrammed in or the robot will need reprogramming. This isn't ideal for many production centres.

We need smart robots. Robots that can adapt to a dynamic environment, learn tasks by themselves and make intelligent decisions based on a high-level understanding of their goals. Reducing the barrier to entry and simplifying the integration process would unlock automation's potential for lower volume production - whether for SMEs or the ever-hyped mass customisation. Not only that, it will allow robotics to easily adapt to minor design changes without causing major issues (see Tesla’s Snafu).

Luckily, the status quo is shifting. Remix is focused on developing highly flexible systems, and many robotics manufacturers are starting to include cognition as a standard, see Neura, Right Hand Robotics and many more.

But what does it mean to be smart?

A recent paper summarises the different strategies for controlling smart robots, laying out the current state of the art and its limitations.

Smart industrial robots follow the principle of “see-think-act” that was originally utilised in the field of mobile robotics -

• See - Sense the environment and extract the relevant info

• Think - Use ‘knowledge’ — high-level instructions and understanding of the context — to plan & adapt based on the environment

• Act - Convert the plan to a path and execute

The paper focuses primarily on the challenges of picking and placing an object which is still surprisingly difficult. There are three major control strategies -

• Deep Learning (DL) - The robot is taught using a set of training data and then applies that learning to real-life situations. In industry, this is often applied using machine vision to determine the object’s pose in order to grasp or manipulate it. Its a fairly common to see simple applications of this technology, but it can get very complicated. For example, random bin picking is seen as one of the industry’s biggest challenges. Issues really start if the product is unknown before entering, or is stacked, tangled, overlapping or in variable environments.

• Deep Reinforcement Learning (DRL) - The robot is taught dynamically by interacting with the environment. It attempts the task and adjusts its actions based on continuous feedback to maximize a reward function. Feedback can include joint position, motor speeds, 3D camera views and a lot more. This approach is currently receiving a lot of interest, despite not yet being widely used in the industry. DRL mimics how a child learns and has the greatest potential for creating a smart system able to flexibly deal with a variable environment. We previously discussed how Deepmind was championing this approach in the search for general robots.

• Imitation Learning - The opposite of DRL. The robot is taught by observing an “expert” and learns about tasks from demonstrations. This approach can have a much lower skill barrier than robot programming as robots can be taught using teleoperation, kinaesthetic teaching, motion capture, virtual/augmented reality and video demonstrations. The key limitation is that the systems are not truly flexible — they are still programmed for a specific task, and can’t deal with much variability.

80% of the time spent on AI projects is spent on the collection and processing of training data. This is even worse if implementing DL and DRL in manufacturing due to the complexity of parts, the impacts of design changes and the environment's uncontrolled nature. The paper suggests that using virtual environments may reduce these challenges. Simulation provides a safe and fully controlled testing environment for DRL and can also be used to generate synthetic pre-labelled data for DL. However, the difference between simulations and real-life reduces a solution’s accuracy when deployed in the field. This explains why Deepmind and OpenAI are using a mix of simulation, real-world testing and imitation to train their systems.

There is still a long way to go before these solutions are robust enough for industrial use but its got us excited.

Video

Stealing jobs? No AI is here to steal our fun.

An engineer has developed a Monopoly AI to better understand the winning strategies - it's all about maximising luck apparently. Highlights include -

• Why you should love trains and Mayfair

• How altruism evolved and was then quickly beaten out of the AI

• Why humans are actually pretty good at the game

As if monopoly wasn't competitive enough… now we have to compete with the ‘Monopoly Gods’

Tweet of the Week

See last week's Bulletin to learn why Tesla built this giant machine