[U47] Rethinking Robotics: the next growth horizon Will. Be. Epic.
Convergence of multiple technologies unlocks automating many tasks no one wants to do anyway
Dear Readers,
In this Update for Product | Strategy | Innovation I will discuss exponential advancements in general purpose, autonomous robots from companies like Amazon, Hyundai, Tesla and Agility Robotics.
Labor shortages in logistics, warehouse fulfillment, manufacturing and other key industries limit economic growth. Hello inflation! Fleets of autonomous robots operating 16 hours per day, 7 days per week will tackle unsafe, boring, repetitive tasks that are hard to fill and maintain with human labor.
These robots require energy and maintenance to operate, but these are far cheaper than the medical care for work-related injuries and ongoing training for each and every worker.
The economic impact for this next growth horizon in robotics will be significant if deployments are sustainable and continue to scale with ongoing innovation over time.
In this Update, I will cover:
BACKGROUND
Amazon Robotics and Proteus robot
Agility Robotics Digit robot and RoboFab
Tesla Bot (aka Optimus robot project)
SOME FINAL THOUGHTS
The leading edge for scaled deployments of general purpose, autonomous robots from multiple manufacturers should start over the next 12-18 months in industrial settings.
These scaled deployments of autonomous robots are an outcome of factories set up to scale their production in parallel to the rapid pace of innovation for the mechanical systems, computer vision, machine learning and process automation that enable these robots to duplicate the work humans don’t want to do.
Amazon, Hyundai and Tesla are not only leading autonomous robot technology innovation. They also have adequate scale to deploy the technology within their own operations to accelerate advancing the technology as early adopters.
Amazon recently announced they have reached 750,000 robots deployed across their global operations. Amazon has started deploying autonomous robots, too.
BACKGROUND
The era of robotics started in the 1950s and 1960s with stationary, reprogrammable, industrial robots. Advances have historically been attributed to specialization to optimize the technology around a unique process.
The convergence of multiple technologies enables a new growth horizon for autonomous robots in industrial settings. Computer vision using advanced machine learning is a primary driver for this convergence.
Instead of programming every movement required to automate a task, advances in machine learning allows robots with the required system attributes to learn, duplicate, scale and improve tasks performed by their human counterparts.
More advanced robot designs will perform a wider range of tasks. This will help to scale general purpose mechanical systems over more specialized technology. This also helps advance the robotics field because the key driver to reduce cost is scaling unit production.
There is an advantage to producing millions of robots with the same mechanical design and can plearn to do many tasks versus different mechanical designs for a limited number of tasks.
Machine learning has improved exponentially over the last decade. This is due to deep learning with multi-layer neural networks for computer vision applications using very large datasets with billions of images for training.
High-performance computing is required to train such a neural network model, but the resulting model can be easily copied and deployed in the field across many robots.
Multiple video streams feed machine learning systems to control the mechanical systems that automate the assigned work. A fleet of robots doing the same task in parallel can upload exceptions and edge cases to train the neural network to learn and improve faster than a robot operating on its own. Scale matters.
Moving materials and goods within a facility may only need wheels as a primary mechanical component to automate work. This is the basis for many Amazon Robotics designs used in Amazon facilities. Proteus is their autonomous robot that can work among its human counterparts. Proteus can transport carts with boxes from one location to another location and stop if someone or other equipment crosses its path to avoid a collision.
Other tasks may require a more general purpose, bipedal human form with robotic arms and fingers to manipulate objects. This could include selecting a specific bin on a shelf, lifting the bin from the shelf and then carrying the bin to another location. This is the basis for Agility Robotics Digit robot.
Tesla announced its plans for a humanoid robot at an Artificial Intelligence event they hosted in August 2021 to recruit top engineering talent. There was excitement over the opportunity, but it was also assumed this was a side project that would not require many resources to prototype and advance the technology in the near-term.
Less than 6 months later, on Tesla’s Q4 2021 earnings call, Tesla announced its Tesla Bot was the highest priority for product development in 2022 across the whole organization. This caught most people outside the company by surprise, but demonstrated Tesla’s commitment to advance and scale autonomous robotics beyond operating and navigating vehicles.
By mid-2023, Tesla stated it had only produced a half-dozen Tesla Bots, but these early versions allowed Tesla to identify and refine the product requirements, technical specifications and product roadmap to scale unit production.
One key insight was commercially available mechanical actuators needed to control many key movements were inferior to the technical specifications Tesla required. This led Tesla to design, develop and scale production of their own actuators with selected vendors. These new actuators should be available in Q4 2023.
Boston Dynamics & Hyundai
An early pioneer for advanced robotics is Boston Dynamics. The company was founded and spun out of MIT in 1992. Early robots from Boston Dynamics were inspired by animals. BigDog, LittleDog, and Cheetah were robots developed out of this inspiration.
BigDog was a project funded by DARPA. Boston Dynamics created a robotic pack mule in 2005 to accompany soldiers in terrain too rough for vehicles. AlphaDog replaced BigDog for military operations.
Boston Dynamics signed a pledge in October 2022 stating it would not support weaponizing any of its robotic innovation. This pledge has also been signed by 5 other robotics companies including Agility Robotics covered later in this Update.
Boston Dynamics has gone through multiple iterations of ownership over the years. They were acquired by Google X in 2013 and then sold to SoftBank Group in 2017. Hyundai Motor Group acquired a reported 80% of Boston Dynamics from SoftBank. This transaction was completed mid-2021. The latter was an important development since Hyundai can also be an end-user of the technology in its vehicle factories to provide early adoption and validate designs and commercial assumptions.
Spot is Boston Dynamics primary commercial autonomous robot. The primary use case for Spot is inspection in industrial settings with 360 degree perception and monitoring capability. Spot is capable of full autonomy including recharging, navigating its environment, inspecting, uploading data and reporting anomalies without human intervention. There approximately 1,000 Spot robots deployed in the field with various customers. Price and the ongoing cost to operate Spot are limiting factors.
Stretch is Boston Dynamics mobile robot for the warehouse to automate unloading cases from trailers and containers at the inbound dock. This is a key use since the fixed cost to implement Stretch can be absorbed across many units handled within a warehouse over time. Stretch can also be used to build orders by loading cases.
Atlas is Boston Dynamics research platform to explore and innovate robotics in the humanoid form factor. Videos of Atlas have helped inspire many people to reimagine robotics and robot capabilities. However, Atlas in its current form is cost-prohibitive for scaled commercial deployments. We will cover more commercially viable options in following sections.
1. Amazon Robotics and Proteus robot
Amazon fulfillment centers are core to its Marketplace and Amazon Prime. To accelerate order fulfillment for 2-day shipping on selected products, Amazon needed to transform its operations.
Kiva Systems in the Greater Boston area was identified as a technology venture with core technology in advanced robotics. Amazon acquired Kiva Systems in 2012 for a reported $775 million in cash. Amazon retained Kiva’s operations in Greater Boston and eventually renamed those operations Amazon Robotics.
The original Kiva robots are deployed within large designated zones at Amazon fulfillment centers to move tall, 4-sided, yellow storage pods containing inventory within the same floor of a warehouse.
This inventory and Kiva robots occupy the majority of the floorspace within an Amazon fulfillment center. The perimeter to the designated zone is used by human operators to either stow a unique product from inbound logistics into inventory or pick a unique product from inventory to fill an order for outbound logistics.
The Kiva robots are used to bring a designated yellow storage pod with a specific product to a specific station so a human operator can complete the designated task.
Employees and contractors are prohibited from entering the designated zone where the Kiva robots operate since the robots do not include features to prevent collision with an unexpected obstacle.
Computers operate the Kiva robots by knowing where all the Kiva robots are located in real-time with extreme precision.
Technicians must wear special technology that emits a signal to immobilize a Kiva robot within a specified radius if the technician must enter the Kiva robot zone to make a repair or clean the floor while the Kiva robots are operating.
This means the Kiva robots and their successors like Hercules and Pegasus cannot operate among human counterparts and other equipment like forklifts.
The Proteus robot is an Amazon Robotics innovation to realize an autonomous robot that can work among its human counterparts. Common tasks include moving carts of stacked cases and totes around the inbound and outbound loading docks.
Proteus can transport these carts from one location to another location and stop if someone or other equipment cross its path to avoid a collision. Amazon Robotics is also working on other designs to automate specific tasks for their human counterparts.
2. Agility Robotics Digit robot and RoboFab
The company Agility Robotics was founded in 2015 to focus on multi-purpose, human-centric robots. Use cases include moving totes & packages, unloading trailers, and last mile delivery.
Offices were originally in Corvallis, Oregon but have expanded over time to include Palo Alto California and Pittsburgh, Pennsylvania. The latter is likely a result of a strong robotics program at Carnegie Mellon in Pittsburgh.
Digit is the Agility Robotics humanoid robot designed to walk forward, backward, side-to-side, turn in place, walk up and down inclines and across unstructured terrain like grass, rocks and curbs.
Digit works 16 out of 24 hours for at least 2 FTE equivalents and can autonomously dock and recharge itself. Digit pauses if it senses a person or object in its path and can then navigate around the obstacle.
Agility Robotics was one of 6 start-up ventures awarded initial funding in April 2022 from Amazon’s $1 billion venture investment program called the Amazon Industrial Innovation Fund.
This fund targets ventures solving the toughest problems across customer fulfillment operations, logistics, and supply chain solutions.
Amazon operations can provide feedback on product requirements and specifications from their subject matter experts to create additional value beyond the fund investment.
Amazon can also deploy technology within its fulfillment centers and other potential operations to verify and validate commercial assumptions for specific use cases.
Agility Robotics has also recently opened the world’s first humanoid robot factory the company calls RoboFab based in Salem, Oregon.
Annual production at RoboFab is planned to reach 10,000 units per year and Digit will work alongside its human counterparts to move totes around the facility to aid production. Digit may eventually take on simple tasks to assist specific process paths within robot production.
Looking at the return on human labor, if each Digit robot produced at the RoboFab provides 2 FTEs to their end-user customer and 100 human FTEs are required plus additional Digit-FTEs to maximize output, then those 100 human FTEs annually would yield 20,000 FTEs of work capacity annually from each cohort of 10,000 Digit robots produced.
3. Tesla Bot (aka Optimus robot project)
A key area of innovation for Tesla is autonomous navigation for its electric vehicles. This includes multiple external cameras onboard the vehicles to acquire continuous 360 degree video, custom hardware to process multiple video streams and end-to-end neural networks to control the vehicle through an inference engine.
These vehicles are equipped to upload edge cases to the Tesla cloud when a human operator overrides autonomous navigation. These edge cases are then verified before they are added to the expanding library of images used to train future versions of the neural network.
Tesla has also deployed its own supercomputing platform to train neural networks for computer vision. Once a neural network is determined to provide a step-change in performance, it can be deployed to the Tesla fleet of vehicles to start the cycle all over again.
Tesla also sees autonomous navigation as a key advantage to improve overall margins for electric vehicles with high-margin subscription services. This is the core reason Tesla is investing billions of dollars to advance its computer vision technology.
Electric vehicles equipped with autonomous navigation also enable the robotaxi use case. This is a specialized use case for mobile, autonomous robotics at Tesla.
As Tesla gained confidence in its computer vision and advanced battery technology, it probably questioned where it could repurpose this technology stack for other applications.
Labor is a significant input cost for Tesla. Some tasks are highly specialized, but other tasks are more routine and low hanging fruit for a general-purpose, autonomous humanoid robot.
As Tesla scales its gigafactory footprint around the globe, its need for labor only increases. This also increases the need for general purpose, autonomous humanoid robots to take on a growing list of unsafe, boring, routine tasks over time.
Tesla announced its humanoid robot concept just over 2 years ago in August 2021. Tesla unveiled an early prototype shown in the video above in 2022 to reveal more details regarding the design and specifications. The Tesla Bot is designed to operate 16 out of every 24 hours 7 days a week. This is equivalent to 112 hours per week or 2.8 FTEs.
The latest update video on Tesla Bot was provided in September 2023 and showcased the humanoid robot autonomously sorting blue and green blocks, balancing for a yoga pose and stretching to demonstrate its agility. The Tesla Bot is trained with videos based on what it would see if was performing the designated task.
The accelerated pace of innovation for Tesla Bot is likely an outcome of the significant progress Tesla has made over the last year with computer vision using end-to-end neural networks. This results in photons entering the cameras, hardware processing the video streams and producing the output control for mechanical systems in either an electric vehicle or humanoid robot. As the core computer vision technology continues to evolve it can be repurposed across multiple products at increasing scale.
Tesla likely prioritized humanoid robot development over its vehicles and energy products because of the business opportunity it provides.
Tesla can repurpose limited space within each of its gigafactories to manufacture the Tesla Bot.
Tesla Bots produced onsite can be used at the same site to start automating unsafe, boring, routine tasks. Each Tesla Bot can provide up to 2.8 FTEs for work.
Over time, Tesla Bots can take on many of the tasks to build new Tesla Bots. This reduces the human labor equivalent to produce each Tesla Bot.
With Tesla’s global supply chain, scale and internal needs to produce robots, vehicles, factories, stationary batteries, install solar roofs and build rockets for sister company SpaceX, Tesla can justify building hundreds of thousands Tesla Bots for its own use.
As Tesla develops and improves the Tesla Bot for its own use at scale within the same facility where they are produced, Tesla will also partner with other companies to provide labor-as-a-service with the Tesla Bot. Unit costs will continue to drop as unit production scales. This improves the unit economics for these services over time.
Tesla has stated its valuation in the future will be based on Tesla Bot and adjacent technologies instead of electric vehicles or its energy business. If Tesla can source the new actuator technology it needs in Q4 2023 to meet its specifications for precise movements, this opens up the opportunity to start scaling production for alpha deployments for selected routine tasks over the next 12-18 months.
Fleets of Tesla Bots doing the same task will accelerate learning and improve training to do this task better over time. This also opens up the opportunity to rapidly deploy Tesla Bots to do this same task across all Tesla facilities. I can see 1000s of Tesla Bots deployed across Tesla and SpaceX facilities before the end of 2025 and 1 million Tesla Bots produced by 2030.
SOME FINAL THOUGHTS
The field of robotics has evolved over the last 70 years with step-changes in technology improving applications. Machine learning has provided one of those step-changes over the last decade to scale robotics for warehouse fulfillment, industrial inspections and autonomous navigation in vehicles. As the neural networks become more end-to-end without the need for code programmed by software engineers, the pace of innovation will continue to accelerate.
General-purpose, autonomous robots trained and operated with end-to-end neural networks will provide a step-change for robotics that will likely surpass all robotic innovation since the 1960s once adequate scale is reached to drive significant value creation. Boston Dynamics, Amazon, Agility Robotics, Tesla and other companies will contribute to that step change.
This step-change for robotics at scale can also provide a step-change for humans if the cost of robot labor continues to decrease over time while sustainable energy also scales to lower the cost of energy. Labor and Energy are two primary inputs that can limit economic growth if supply becomes constrained. Economic output cannot be infinite, but it can grow by an order of magnitude two times faster to accelerate the value created.
There are societal issues, too, if robots start to displace many employees from jobs they want. But we can start with automating the many unsafe, boring, routine jobs we do not want. And then over time we need robots to complement human labor to expand our capabilities. Labor for many jobs will need to continuously up-skill and then backfill more of what they use to do with a team of robots. That is easy to say, but much harder to do. However, Darwin is a great mentor regarding the need to adapt.
Best,
Stephen
I’m long AMZN and TSLA mentioned in this Update. Nothing in this Update is intended to serve as financial advice. Do your own research. The opinions and views expressed in this newsletter are those of the author. They do not purport to reflect the opinions, views or policies of any other organization, company or employer.