Inspiration
We were inspired by a gap between what warehouse automation promises and what workers actually experience. While systems like Vulcan are designed to improve efficiency, the reality is that fulfillment center associates still face significant physical strain and constant workflow disruptions. From our research, we found that associates experience 15–20 robot failures per shift, forcing them to repeatedly stop their work, intervene, and reset the system. At the same time, workers are climbing ladders and bending for low-bin picks throughout 10-hour shifts, contributing to an injury rate of 7.3 per 100 workers, which is significantly higher than the industry benchmark of 2.3 per 100. Despite having 750,000 deployed robots, associates were still climbing ladders repeatedly during 10-hour shifts and being interrupted 15 to 20 times per shift to manually correct Vulcan's failed picks. We saw a robot that existed but wasn't solving the full problem, and we wanted to fix that through software.
What it does
How we enhanced Vulcan AI is an Intelligent Human-Robot Collaboration System that sits on top of Amazon's existing Vulcan robot. It predicts when Vulcan will fail a pick before it even attempts it, then dynamically routes that task to the right person or robot in real time. Associates only receive mid-height picks, eliminating all ladder use. At its core, the system uses machine learning to analyze item characteristics such as packaging type, size, weight, and historical pick success rates, to determine the probability of a successful robotic pick. Based on this prediction: High-confidence tasks are assigned to the robot, High-risk or complex tasks are routed to humans
How I built it
We built it as a software-first system using three core components: a machine learning failure prediction engine that analyzes item characteristics before each pick, a dynamic task routing system that integrates with Amazon's existing warehouse management system, and an associate-facing touchscreen dashboard mounted at each station. No hardware modifications to Vulcan were required.
Challenges I ran into
One of the biggest challenges was working within real-world operational constraints. The system depends on real-time WMS integration, so even small delays can create bottlenecks on the warehouse floor. We had to design fallback mechanisms to ensure the system could still function reliably during outages or latency issues. Finally, balancing efficiency with safety was critical. The system needed to optimize throughput while strictly enforcing ergonomic constraints, with zero room for error.
Accomplishments that I'm proud of
We are really proud that our solution doesn’t just improve efficiency, it also makes the job better and safer for the people doing it every day. We’re reducing robot failures, improving overall speed, and increasing accuracy, while also minimizing physical strain by eliminating tasks like unnecessary ladder use and frequent manual interruptions. What matters most to me is that we didn’t just focus on fixing the robot, we considered the system as a whole. Our goal was to create a setup where humans and robots work together more effectively, rather than getting in each other’s way. That perspective is what makes this solution both practical and scalable, with the potential to create real impact.
What I learned
We learned that the real problem was not the robot it was the system around the robot. Associates were not failing because Vulcan was bad hardware. They were failing because no intelligence existed to decide who should do what before the robot made a mistake. We also learned that FC associates are reached through internal and physical channels, not traditional marketing the A to Z app and the station touchscreen are the only channels that matter for this user.
What's next for Vulcan AI
The immediate next step is a single-zone pilot at Spokane in June 2026 with a small team of associates and a couple of Vulcan units to test the system in a real operating environment and refine it based on on-the-ground feedback. If that goes well, the plan is to expand to Hamburg, followed by a full fulfillment center rollout by October 2026, and then scale to additional sites across the U.S. and Europe by the end of the year. In the longer term, the goal is to connect learnings across facilities so improvements in one site help strengthen performance everywhere. We also want to evolve the system into something more intuitive, including voice-based, hands-free interaction for associates on the floor to make daily operations even smoother.
Built With
- canva
- figma