Innovation has been embedded in Picnic’s DNA since our inception. From reinventing traditional grocery delivery to developing fully automated warehouses, our focus has always been on finding ways to boost efficiency and improve the satisfaction of our customers. One significant step forward on this automation journey is the integration of robotic arms into our order-picking processes. This post explores in detail why, how, and where we’re integrating these robotic arms, as well as the challenges we face and our strategy for future growth.
Robots, Robots & More Robots
People often associate the word “robot” only with robotic arms or self-standing humanoids. However, the Cambridge Dictionary defines a robot as “a machine controlled by a computer that is used to perform jobs automatically”. Under this broad definition, robots have played a crucial role in Picnic’s warehouse operations for years.
For instance, our automated fulfilment centres in The Netherlands and Germany operate thousands of robotic systems to ensure crates of every kind move seamlessly through our operations. A recent article from the German press even highlighted that our warehouse in Oberhausen (Germany) is operated by approximately 1,500 robots, showing our long-standing experience with robotic automation.
Moreover, we already have four robotic arms in operation at our warehouse in Dordrecht, supporting the dispatching process by loading full crates into frames. These frames are then placed in our electric vehicles for delivery to customers, showcasing our continuous push towards automation.
Robots and Picking
Despite the high level of automation within Picnic’s operations, several processes still rely heavily on human labour:
- Decanting: The transfer of products from pallets into storage crates.
- Rework: Handling exceptions and anomalies within orders and crates.
- Picking: Selecting and preparing items for customer orders.
Among these, picking is particularly labour-intensive as well as critical in all our operations as it is how orders are fulfilled. Our Goods-To-Person (GTP) pick stations are designed to minimize physical strain, routing both the stock crate (containing articles) and the customer order crate directly to the picking point where an operator (shopper) stands. Both crates are within arm’s reach, reducing the picking action to just transferring items from the stock crate on the right to a designated compartment in the customer order crate on the left. This environment provides the ideal setup to introduce robotic arms, as they can be fixed to the station.
Goals and Objectives
The robotic picking initiative was launched with several goals in mind, ensuring that every step aligns with our operational and customer-focused priorities:
- Support Flexible Operations: Enable 24/7 operational capacity, with robots handling picking during off-peak hours such as nights, weekends and lunch breaks.
- Maintain Quality Standards: Ensure that product handling meets Picnic’s high-quality standards, ensuring that the orders are picked with the highest customer satisfaction.
- Generate Actionable Data: Gather detailed performance metrics from robotic operations to inform future automation decisions.
We are approaching the initial stages of this initiative with an experimental mindset, enabling us to fully understand both the capabilities and limitations of the robots.
- Product selection: Identify which products the robot can reliably pick
- Volume stress: Determine the order load a robot can handle during a standard shift.
- Fill-rate capacity: Assess the maximum crate volume the robot can accurately pick into.
- Error recovery scenarios: Observe how the robot responds to issues such as broken products, missing items, or incorrect products in the stock crate.
Each experiment was properly tracked, enabling us to quantify performance across key metrics, including: Pick accuracy and success rate per product, average pick time per item, and downtime and intervention frequency.
Challenges with Robotic Picking
Robotic systems offer considerable benefits, but several limitations persist, especially within picking tasks:
Complexity of Items: They can handle a wide variety of products, but not all of them. Items with irregular shapes (like bananas), fragile ones (such as eggs), or expensive products (like champagne) require careful handling that only very specialized robots can match. Therefore, we restrict robotic-picking to products suited for automated handling.
Speed and Adaptability: Nothing is faster than our shoppers — seriously! Picking at our Goods-To-Person stations involves moving items smoothly from one crate to another, an action currently performed by humans with unmatched efficiency. Modern robots simply can’t replicate this speed and fluidity yet.
Fill-Rate Constraints: Picking into full crates isn’t challenging for humans. A human can easily rearrange items within a crate to maximize space and fit more products. Robots, however, require predefined positioning, restricting their capacity to maximize space utilization.
To address these constraints, we introduced several mechanisms:
Product whitelisting, a mechanism that specifies which items can be picked by robots, respecting fragility constraints. For example, if the robot can handle packaged snacks but not glass bottles, orders containing both snacks and bottles should never start at the robot station. You wouldn’t want your snacks damaged by a heavy bottle.
Fill Rate limit: We restrict the number of products a robot can place into a crate. This restriction is based on both the total volume of the crate and the volume of each individual item being placed. If an order exceeds the robot’s fill-rate limit, the robot picks only part of the order, and the remainder is later completed by a human at another station.
Integrating Robots into Picnic’s Technical Infrastructure
We took a pragmatic approach for Incorporating third-party robotic systems into our existing infrastructure: having the robots emulate the interactions of our stations UI client, consuming the API designed for our stations client to ensure seamless integration.
Understanding WCF
We call the user interface application running at each picking station the Warehouse Client Fixed (WCF). The client is rendered at every station and directly guides our shoppers by indicating what actions they need to take for the orders in front of them. These actions include picking items, counting stock, resolving issues, among others.
The Warehouse Control System (WCS) communicates these actions via messages handled by an intermediate layer that converts them into Server-Sent Events (SSE) transmitted through event streams, which are established once a station is connected and a user has logged in. These event streams continuously update the UI client with new instructions. Operators confirm completed actions through the UI, which then calls corresponding endpoints on the WCS.
Our integration strategy for robots involves allowing them to directly consume messages from WCS and confirm picks through our existing API endpoints. This enables faster integration and reduces complexity on our backend. Robot vendors, however, remain responsible for integrating their software with our existing APIs.
Operational Workflow Adjustments
Now, the robot is deployed, can receive instructions, and confirm picks. Seems like everything is ready, right? Not quite. We still need to address previously mentioned limitations:
Selective Order Allocation: Robot stations receive filtered orders tailored specifically to robotic capabilities, unlike human-operated stations, which handle a wider variety of tasks continuously.
Partial Picking and Forwarding: For orders exceeding robotic capabilities, either due to fill-rate limitations or inclusion of non-whitelisted items, the robot partially fulfils the order, with the remainder being forwarded to a human-operated station. Our WCS carefully tracks these intermediate states to ensure smooth handovers.
Exclusion of Ancillary Tasks: Tasks such as inventory counting and quality checks, typically conducted at manual stations, are currently disabled at robotic stations due to present limitations in robotic capabilities.
Ensuring Reliability Through Monitoring and Recovery Systems
Robotic systems must operate reliably and predictably. Thus, robust monitoring and fall-back mechanisms were essential additions:
Order Line Marking: To evaluate the robot’s ability to handle items outside of the white-listed ones, we’ve introduced a feature called Order Line Marking. This lets us tag specific order lines (Often from internal orders, such as those used to supply canteens in our warehouses) for close supervision. These tagged lines help us assess how well the robot performs on edge-case products, and where its current limitations still lie.
Analyst UI Driver: Although the robot interacts directly with our backend APIs, we’ve also deployed a lightweight version of our picking station UI for on-site analysts. This client displays the robot’s intended actions in real time (e.g., “pick 3 bananas into crate compartment A”). If the robot gets stuck or fails to act, an operator can manually intervene, recording what action was taken and, if necessary, stopping further work on the order.
Real-time Data Monitoring: We are a fully data-driven company. Therefore, we provide comprehensive tools that enable analysts to determine whether the robots are performing optimally or if we can potentially increase the workload or expand the variety of products picked by robots. Analysts can even monitor in real-time, from their laptops, if a robot is experiencing issues.
Our WCS reports so-called Analytical Events to our Data Warehouse for every action occurring at each station. This data provides valuable insights, such as the picking speed, the number of orders the robot fulfils per hour, and whether a robot is encountering difficulties with specific products or not picking at all. The data in-feed dashboards in Grafana for real-time supervision.
Vendors also provide a SCADA (Supervisory Control and Data Acquisition) application that enables real-time monitoring of the machinery’s state. With this combined set of tools, we can evaluate whether a robot is reliable in terms of hardware (i.e., fully operational during warehouse activities) and business objectives (i.e., meeting our expectations regarding picking rates and the range of products handled).
Scaling Robotic Picking Across Picnic
The end goal of our robotic picking initiative is a gradual roll-out across all GTP stations.
Additional operational enhancements will likely include optimization based on actual robot performance data, expanded product white-lists informed by real-world results, and refined integration between human-operated-stations and robotic tasks.
Conclusion: A Human-Centric Approach to Automation
So, will robots replace our shoppers? Absolutely not. As mentioned, it isn’t our goal to replace them either, but rather to use robots to boost our warehouse’s performance. Shoppers remain at the core of our warehouse operations, with robots complementing their efforts.
The results we’ve seen so far from our ongoing initiative are promising. They reinforce our belief that automation, done right, can improve productivity and help us to better serve our customers. At Picnic, we’re taking big steps to integrate robotics, always guided by the principle that technology should serve people, not the other way around.
Automating the Pick: How Picnic is Integrating Robots in Their Order-Picking Process was originally published in Picnic Engineering on Medium, where people are continuing the conversation by highlighting and responding to this story.
