Solving The Last-Mile Delivery Problem

Autonomous roadside delivery robots may increase operational efficiency, accuracy, and customer satisfaction.


Retailers are deploying robots to cut costs and improve efficiency, opening new opportunities for chipmakers as well as a host of new challenges.

Key to this strategy are autonomous roadside delivery robots (ARDRs). Retailers have been facing razor-thin profit margins for years and have turned their sights to increasing operational efficiency to stay competitive. Solving the last-mile delivery problem within the supply chain is an important piece of the efficiency puzzle, regardless of whether distribution stems from warehouse or a grocery store. Intelligent robots potentially can increase efficiency, accuracy, and customer satisfaction, and they can make deliveries in less time with lower operational and indirect costs.

ARK Invest estimates that robots could deliver food for a cost of about 6 cents per mile, a 20X cost-saving compared with delivery by humans. Although using larger delivery robots costs a little bit more — about 40 cents per mile — that is still one-sixth the cost of taking a personal trip to the grocery store.

This is reflected in the growth projections for the autonomous delivery market, which Technavio predicts will grow to $24.8 billion by 2026 at a CAGR of 19.85%. The last mile opportunity is even larger. Straits Research pegs last-mile delivery market at $123 billion by 2030.

Amit Kumar, director of product management and marketing for Tensilica Vision, radar and lidar DSPs at Cadence observed that the biggest benefit for automation delivery services is efficiency as robots can work 24/7 and bring cost down significantly when deployed at scale. “Most delivery robots are electrically powered, hence they help reduce greenhouse emissions and are environmentally safe. They are also capable of doing contactless deliveries, which in the case of a pandemic era are preferred over human delivery services.”

“This technology is enabling last mile delivery (LMD),” Kumar said, “which means that a human driver take the packages to a specific location (or hub), and from there these robots will take over delivering goods to customers whether institutional (offices) or retail (homes). Efforts are underway to use robots for mid-mile deliveries (MMDs), which are the longer routes. A delivery robot is essentially a Level 4/5 vehicle, which is fully automated and/or monitored by a remote operator who is looking at the fleet of robots and ensures smooth operations.”

The ARDR market has attracted a slew of investors. Today, there are many ARDR manufacturers, including Starship, Nuro, Udelv, Coco Robotics (a.k.a. Cyan), Kiwibot, Eliport, and others.

Starship Technologies has offered Level 4 autonomous robots for sidewalk delivery since 2018. Today, its fleet of two-thousand robots uses a combination of GPS and sensor fusion to navigate. The technologies include 12 cameras, ultrasonic sensors, radar, neural networks, and AI to detect obstacles, enabling the robots to map the environment to one-inch accuracy.

Starship has created different models, based on what works best in each community, and is currently providing autonomous deliveries in nearly 50 service areas, ranging from college campuses to neighborhoods. In the United States, most deliveries are done on college campuses, while neighborhoods are the leading application in other countries. A new delivery as a service (DaaS) launched recently with a robot fleet available for rent. Starship also formed a partnership with Grubhub in the U.S. and S-Group in Finland. In April 2023, the company announced the total robot miles traveled, which was six times more than the milestones announced earlier this year by Waymo and Cruise.

Fig. 1: The autonomous robots from Starship is equipped with cameras, ultrasonic sensors, radar, neural networks, and AI to detect obstacles, enabling the robots to map the environment to one-inch accuracy. Source: Starship

Fig. 1: The autonomous robots from Starship is equipped with cameras, ultrasonic sensors, radar, neural networks, and AI to detect obstacles, enabling the robots to map the environment to one-inch accuracy. Source: Starship

Nuro, founded in 2016, has three different ARDR models, including one that measures 10.5′ x 7′ with a maximum speed of 45 miles an hour. This ARDR is equipped with long- and short-range sensors, thermal imaging cameras, radar, and audio/microphone sensors. Additionally, it has an external airbag and sound generator for pedestrian protection and can carry a maximum load up to 500 pounds.

Fig. 2: Nuro’s delivery robot can carry a maximum load up to 500 pounds, with a maximum speed of 45 mph. Source: Nuro

Fig. 2: Nuro’s delivery robot can carry a maximum load up to 500 pounds, with a maximum speed of 45 mph. Source: Nuro

Fig. 2: Nuro’s delivery robot can carry a maximum load up to 500 pounds, with a maximum speed of 45 mph. Source: Nuro

Nuro, in partnership with Uber, will automate delivery for Uber Eats. Consumers in Houston, Texas and Mountain View, California, can order meals and goods with Nuro’s zero-occupant autonomous delivery vehicles. The companies signed a 10-year contract. If the pilot program succeeds, Uber Eats said it will roll out service with no driver. Similar to other ARDR providers, Nuro uses remote monitoring as a safety measure.

The technology behind ARDR
An ARDR is a mini version of a regular autonomous vehicle (AV), with many of the features required by regular AV. Typically, they operate at a lower speed, topping out at approximately 45 mph. In some cases, these autonomous delivery robots are capable of Level 4 autonomy, one level up from the Level 3 autonomous vehicles commercially available today. ARDR has full sensor fusion capability. Additionally, most ARDR manufacturers provide remote monitoring and control to deal with accidents and emergencies.

“At a basic level, delivery robots share a similar sense-plan-act pipeline with highly automated vehicles,” said Guilherme Marshall, director, Automotive Go-To-Market at Arm. “Delivery robots also can benefit from many of the recent advancements in sensor, compute, and simulation technology spurred by automated driving.”

At the same time, the restricted operational design domain and less stringent homologation requirements for delivery robots means it is significantly easier to design and validate compared with full-size vehicles operating on public roads. Nevertheless, there are still plenty of corner cases that need to be considered.

“On one hand, it is interesting to observe that on these small robots — propelled by very modest electric traction systems — compute can become a much larger proportion of the overall power consumption,” Marshall explained. “At scale, it becomes clear that commercial robot operations running on highly efficient compute platforms can gain significant advantage by reducing downtime for recharging.”

According to Cadence’s Kumar, to make the delivery robots work requires a combination of many technologies including:

-Sensor suite includes vision + radar + lidar (sometimes a subset of these) that provides 360-degree visibility and helps a robot perceive the environment.

-Wi-Fi is necessary for updates both to and from the robot. This includes both OTA updates and taking control of robots in an event since their control modules are connected to the central monitoring stations.

-GPS helps a robot navigate the area either via a pre-existing map and making sure it is following the route, or more recent routes where it helps the robot go from Point A to Point B.

-Compute usually involves a system on module (SOM) powered by a GPU or an SoC that is responsible for taking sensor inputs and processing the data, fusing the data, and coupled with the AI engines making the final decision.

-Batteries are a very important aspect of a robot because the denser the battery storage, the more miles a robot can work before returning back to a charging dock.

-Battery management systems are being designed specifically for smaller batteries that go into robots and could be air cooled or water cooled.

-Speed: Usually last mile delivery robots operate at lower speeds. However, there is no standard per se and it depends on the local city/county/state regulations and the type of environment it operates in. On average, it runs between 4 miles ~ 6 miles per hour. Lower speeds also allow them to safely navigate pedestrian areas and also allows them to share a sidewalk in an event they have to get off the road.

-Ease of Design: Though these are no standards that govern delivery robot designs, it becomes a challenge for the OEM to do sensor selection, positioning and placement, battery choices, regulatory approvals, etc. A classic example is functional safety; even though a robot is like an electric vehicle, ISO26262 that used to apply is being changed to ISO22116. This standard is not fully developed for industrial mobile robots; however, the guidelines provide operation assessments to mitigate risks, e.g., emergency stop, collision avoidance, etc. However, the system requirements to operate the robot are more relaxed than an electric vehicle, and so are the design requirements.

AI required
AI is also a key component in ARDR design. And as this technology transitions from startup stage to mainstream, developers will struggle with balancing the requirements of performance, which is heavily dependent on AI as well as a host of other sensors, and the cost of the technology needed to make this all work.

“Delivery robots are in many ways similar to fully autonomous vehicles in both characteristics and design challenges,” said Geoff Tate, CEO of Flex Logix. “Small size means they are more power- and size-constrained. Slow speed means they don’t need to run as many inferences per second. Initially, with lower demand, it means full custom SoCs are not economic (unlike high-volume cars). So for AI, these products in the early years will use off-the-shelf modules from companies like TI and NVIDIA that were developed originally for cars. As volumes grow and SoCs become feasible, they will use AI IP.”

Even though ARDRs operate at a lower speed than autonomous vehicles running on the freeway, reliability and safety are equally important. ARDR can learn from what OEMs have done in the past.

“Autonomous robots are leveraging many of the same technologies that autonomous vehicles are developing,” said Ron Lowman, strategic marketing manager at Synopsys. “Reliability and safety are paramount to autonomous vehicles, and OEMs have put a great deal of effort into developing products meeting those standards.”

Additionally, a lot of the underlying technologies from the semiconductor industry have been developed for the automotive industry for many years.

“Autonomous robots and delivery vehicles for groceries or other goods can take advantage of what has been developed for the automotive industry,” Lowman said. “Each industry has its own specs and mission profiles. Autonomous vehicles are concerned about the safety of the human beings inside and outside of the vehicles for a typical life of a vehicle. A good number of the safety standards, including IEC 61508 and ISO 26262, are being implemented. While fully autonomous robots do not have a human inside, ARDR manufacturers want to make sure humans are safe and that the robots will not damage anything. Therefore, those safety standards and others like them are also applicable throughout the entire supply chain.”

ARDR design challenges
Because many Level 4 autonomous delivery robots use similar autonomous technology to AVs — including AI, sensor fusion, embedded vision, emergency braking systems, remote connectivity, and GPS — optimizing performance and cost is important in product development. So are safety and security.

However, unlike autonomous vehicles, delivery robots interact with many different people. Therefore, robots need built-in intelligence to assure transactions run smoothly. The ability to navigate objects, even moving objects such as joggers and dogs, is a must. So is monitoring the battery level to avoid stranding the ARDR. In some ways, designing ARDR can be more challenging than designing a semi-autonomous vehicle because the ARDR is fully autonomous and has to be aware of its surroundings without human interaction.

Further, as Kumar stated above, the system is designed to prevent robots from running into things that are in front of it, such as pedestrians, pets, bicyclists, vehicles, etc. “GPS usually prevents the robot from getting lost and enables the robot to be recoverable via its last known location in the event its battery runs out. A remote operator can also assist in maneuvering the robot and in the event it gets stuck, can raise an alarm to recover it.”

Additionally, Amol Borkar, director of product management, marketing and business development, Tensilica Vision and AI DSPs at Cadence noted that collision avoidance is a critical aspect of design for safety. “A few years ago, a security robot accidentally ran over a toddler, resulting in injuries to the young one. Such accidents can lead to immediate decommission of the robot and hefty lawsuits. Therefore, as part of the design, such robots are typically equipped with an array of different sensors, e.g., camera, short-range radar, bump sensor, etc. to provide redundancy and allow the combined perception to be better than a single sensor, which could help avoid such accidents in the future.”

ARDR manufacturers must have remote monitoring and control capabilities to rescue the robots whenever there is a problem. Because they operate in public spaces, the robots could be a target for thieves. In some cases, the size makes it possible to simply pick up the units, and proactively preventing this can be a challenge.

Starship said that with multiple cameras installed, the environment could be monitored. This may require AI and/or constant remote monitoring.

“These small robots are being designed to face many adversities in their journeys,” Arm’s Marshall said. “For instance, near the Arm office in Cambridge, it is common to see robots struggling to move past vehicles parked on narrow pavements. Delivery robots must harmonize and even collaborate with road/pavement users in ways that perhaps go beyond automated vehicles. Consider for a moment that some pedestrian crossings only stop vehicles once a button has been pushed by the person waiting to cross the street. In this situation, robots will need to communicate with pedestrians and depend on them for a safe crossing. In addition to remote monitoring, fully remote operation and audio communication are likely to become design requirements.”

Designing autonomous delivery robots also comes with several challenges that must be addressed to ensure their safe and efficient operation.

“Some of the critical design challenges include collision avoidance and object detection, localization and mapping, path planning and navigation, adaptation to changing environments, robustness and fault tolerance, security and privacy, remote monitoring, and supervision,” said Andy Nightingale, vice president of product marketing at Arteris IP. “Addressing these design challenges is essential to ensure autonomous delivery robots’ safe, reliable, and effective performance. Continuous research, development, and testing are vital to refine the designs and algorithms, making them more robust, adaptable, and capable of handling the complexities of real-world delivery scenarios.”

The role of IP in ARDR design
Similar to other autonomous vehicles, semiconductors and IP also play an important role in ARDR designs. CPUs, GPUs, and AI have, and will continue, to play a key role in ARDR, and IP provides important building blocks for future designs.

“Semiconductors and IP play a vital role in this category” Kumar said. “In the past, computer racks were used to make a fleet autonomous. This was not an option for robots due to their size and weight-carrying capabilities. With the advancement of VLSI and reduction in process geometries for SoCs, higher-compute capabilities are now available, making the size of the computer “system on module” very small and efficient. IP such as vision processors, lidar/radar processors and AI engines are at the heart of such semiconductor devices that make it all come together by taking in multi-modal /multi-stream input at the edge for processing and utilizing AI engines running ML models trained on specific use cases to provide intelligence to a machine.”

Borkar agreed semiconductors play a vital role in this. “In the past, off-the-shelf components, CPUs etc. were used to control and navigate robots. However, as the requirements and specifications in this segment have also become more involved and tailor-made, going down this route is not feasible for both performance and efficiency. With the evolution of AI being used in nearly all autonomous applications for perception and other components of the flow, having specialized processors and accelerators for AI targeted at the robotics market also makes sense.”

“The industry already offers a large range of SoCs that can scale to offer suitable solutions for the start of operations in autonomous robots,” Marshall said. “CPUs, GPUs, and ISPs have been designed to provide leading performance per watt as well as functional safety built in from the ground up. These will be key enabling factors to scaling autonomous robot operations with minimal asset downtime and to help create trust in the public.”

Further, semiconductor IP provides the necessary building blocks and components for developing and deploying autonomous delivery robots, which include everything from sensor integration and processing power to communication, memory management, power optimization, security, and design tools.

Nightingale pointed to various design components and building blocks for ARDR, including:

  • Sensor integration: IP provides sensor interface IPs that enable seamless integration of these sensors into the system, allowing the robot to capture, process, and interpret real-time data for navigation, object detection, and obstacle avoidance.
  • Processing power: IP components, such as processors, digital signal processors (DSPs), and accelerators, provide the computational power necessary for the efficient operation of autonomous delivery robots.
  • Communication and networking: IP for communication protocols, such as Ethernet, USB, CAN, and wireless connectivity (e.g., Wi-Fi, Bluetooth, and LTE), facilitate seamless communication between the robot’s subsystems, control systems, and remote monitoring centers.
  • Power management: IPs such as voltage regulators and power controllers, along with low-power design techniques, help optimize power consumption, extend battery life, and enhance overall system energy efficiency.
  • Security and encryption: IP for encryption, authentication, and secure boot mechanisms ensures data integrity, confidentiality, and protection, preventing unauthorized access and tampering.
  • Verification and design tools: IP providers also offer verification IPs (VIPs) and design tools that assist in developing, testing, and verifying autonomous delivery robot systems. These tools help streamline the design process, ensure compliance with industry standards, and improve overall system reliability.

One way to utilize these building blocks is by implementing a network-on-chip (NoC) IP technology into the SoC, which will help optimize communication infrastructure, network interoperability, scalability/modularity, and efficiency, supporting quality of service.

Future trends
It’s tough to accurately project how this market will evolve, given that it’s early days for a last-mile delivery solution, which at this point is an expensive experiment with an uncertain future. Still, investors have poured more than $8 billion into the autonomous delivery market, including both ARDR and drone delivery.

Even though many are jumping in, not everyone is convinced. Two major delivery companies, Amazon and FedEx, have canceled their ARDR programs.

In October 2022, Amazon decided not to continue with its ARDR project. “During our limited field test for Amazon Scout, we worked to create a unique delivery experience, but learned through feedback that there were aspects of the program that weren’t meeting customers’ needs. As a result, we are ending our field tests and reorienting the program,” according to Amazon spokesperson Maya Vautier. Instead, Amazon will focus its efforts primarily on drone delivery.

FedEx launched its delivery robot program, the FedEx Sameday Bot, in 2019, but the company decided to scale back its efforts in 2022 because the program did not yield short-term benefits.

It’s difficult to predict what the last-mile delivery solution will look like five years from now. Future ARDR will come in different shapes and sizes. But this is only one approach. Other concepts are sure to emerge.

For example, Robomart is taking a different approach. Instead of using autonomous robots, the company has created mobile stores operated with a human driver. The company’s slogan is “No cart. No checkout. No hassle.” Consumers will be able to choose a Robomart type and see what it stocks before summoning it with a mobile phone app. By swiping the app to open the door, consumers can take the item, which will be tracked automatically with the receipt printed in seconds once the credit card on file is charged. This is still in an experimental stage. The company claims that the mobile store will arrive within 10 minutes. The obvious question is how many types of mobile units will be required to satisfy consumers’ demands. Additionally, as companies such as Amazon are pushing for drone delivery, how will this impact the development of ARDR?

Going forward, cybersecurity is likely to garner increased attention in this space to ensure that someone doesn’t hack or take control of the robot. “In addition, there is a lot of valuable data that goes to and from the robot while customers access it using their cellphone,” Kumar added.

There are many potential benefits of using intelligence ARDR for the last mile delivery including increased efficiency, accuracy, and customer satisfaction, resulting in reduction of overall operational costs. Many investors and companies are pursuing this multibillion dollar market.

However, there are still many unanswered questions. If hundreds of autonomous roadside delivery robots are running around a neighborhood, it could potentially create a chaotic situation. What types of city ordinances and regulations would be required to direct this ARDR traffic? If an accident occurs in which a pedestrian is hurt, whom will the police or other authorities go after? This will create extra work and bureaucracy for both the city and insurers. How will this potential increase of risks offset the efficiency and cost savings offered by ARDR?

These questions must be answered over time as the search for the last-mile delivery solutions continues.

Leave a Reply

(Note: This name will be displayed publicly)