Pickle is Physical AI

DAE runs the sense-move-manipulate loop thousands of times per shift, in trailers where the scene changes with every pick. It’s the autonomy that lets a Pickle robot work in a 130-degree trailer in summer.
‍
The autonomy stack combines machine vision with a battery of other sensors to see and sense the freight inside trailers, move the robot vehicle, move the robot arm, grasp packages, manipulate them inside the constrained space of a trailer, and place packages without damaging them. It repeats this sense-move- manipulate coordination 1000’s of times per hour and it must deal with an infinite number of complex scenes.

The Dill Autonomy Engine uses generative AI policies (diffusion models) fine-tuned on Pickle’s own warehouse data. These policies turn classical motion plans into precise picks across messy, varied freight.
‍
This enables Pickle to rapidly create custom models for business-specific use cases based on the knowledge encoded in pre-existing models. This approach delivers real value for customers by providing products that just work on day 1 and get better over time.

Pickle’s vision system evaluates pick and path options for every package in milliseconds. Multiple cameras and real-time signal processing handle the scene changes that happen after each pick.
‍
We employ multiple cameras, analog-to-digital conversion and digital signal processing in real- time to manage the complexity. The scenes the systems deal with constantly change. With every package picked and placed, the machine vision is presented with a new scene that unlocks multiple pick and path planning options, but the system can’t sit and think too long because the real world requires picking speed. Our vision system is built to accommodate the real work of unloading trailers as fast as possible.

Pickle’s suction grippers handle packages up to 50 pounds. In a constrained trailer, the robot has to grasp boxes from the front. DAE’s planner picks the grasp and the policy places it.
‍
The constrained physical space of a trailer requires that packages near the ceiling be grasped from the front of the package. Pickle’s suction-based solution excels at picking and placing large heavy packages no matter the location or freight configuration. Gripping methodologies are ultimately bounded by physics and Pickle’s grasping and manipulation capabilities push the boundaries of physics with every pick.

Pickle Robot systems use proven industrial components: KUKA robot arms built for rugged duty, on-board compute, integrated software-controlled lighting. Pickle engineers and tests every custom component end to end.
‍
Pickle Robot systems deliver exceptional performance in non-pristine industrial environments by leveraging proven technology and robust components. The systems rely on durable hardware like Kuka robot arms, built specifically for rugged applications, and integrate commercially available equipment like conveyors to avoid unnecessary reinvention. Pickle designs and rigorously tests all components that aren't commercially available, engineering the entire system and software in tandem to ensure reliable operation at scale. Each unit is self-contained, featuring on-board compute and integrated, software-controlled lighting.

Pickle robots generate system level data that is used to improve system performance via AI learning. They also capture operational data that warehouse managers and inbound supervisors use to run efficient operations in real-time.
‍
Operational metrics are presented to operators via cloud- native dashboards. Metrics include picks per hour (PPH), container/trailer unload times, and information about the package mix. With Pickle’s robust data capture engine running in the background, customer specific dashboards can be configured to monitor and manage single robots, fleets of robots within a site, and fleets of robots across multiple sites.