Ambi Robotics Inc. today unveiled AmbiStack, an overhead palletizing system designed to stack items onto pallets or into containers. The company said it applied its experience with picking and moving boxes and packages to developing the system.
A four-axis gantry robot is at the center of AmbiStack. It sits on top of an infeed conveyor and uses a camera, machine vision, and artificial intelligence to identify items on the conveyor, acquire them, and then stack the items.
The outgoing pallets can be manually moved into and out of the robots’ workspace, similar to how operators interact with gaylords and carts in the existing AmbiSort sortation system. AmbiStack restricts access of the robot to the full pallet while the operator removes a pallet or container from the workspace. Light curtains and floor safety lasers provide safety in the robots’ workspace.
Access to the pallet locations is from the perimeter of the gantry workspace. The initial work envelope of the gantry will be sized to contain six or 12 pallet locations, said Ambi Robotics.
AmbiStack ready to stack items on the fly
AmbiStack is designed to handle both uniform and non-uniform packages coming down the conveyor, according to the Berkeley, Calif.-based company. Its AI technology allows the robot to stack items without prior knowledge of their size, position, or appearance. This advance in “physical AI” adds stacking as a robot skill in AmbiOS, said Ambi.
Powered by an AI vision system based on foundation models, AmbiStack analyzes, tracks, and picks each item while performing quality control checks.
Ambi Robotics is using reinforcement learning to train the AI model for the palletizing robot, explained Jeff Mahler, co-founder and chief technology officer of the company.
The reinforcement learning approach involves defining a reward function that incentivizes the robot to build stable, dense pallets. This reward function is based on models of factors such as item stability, friction, and crushability.
Reinforcement learning trained in simulation
AmbiStack applies reinforcement learning to bridge the simulation-to-reality (Sim2Real) gap and enable robots to stack any item they receive from Day 1 without the need for pre-programming, said Ambi Robotics.
“Within the simulation, we make sure that if we say that a stack is good, it would be good in the real world,” noted Mahler. “So we need models of the stability of the stack, how items lay on top of each other, and the friction of how those items could slide on each other.”
“Also important is crushability,” he said. “Is an item going to get crushed when there’s too much weight on top of it for the material that it has?”
Ambi Robotics released the latest generation of its foundation model for warehouse robotics called PRIME-1 earlier this month. The new generative AI model uses simulation to create the data for the model.
“We’ve built a lot of learnings that we have from traditional robotics,” said Mahler. “For reinforcement learning, the way to do that is through a reward function. So the machine’s only going to get rewarded if it builds essentially a stack that stays stable up till the end, and then it gets essentially a reward for how much of the space it utilizes or the density of that stack.”
- “We always want to fit more items into a given space, and this is essentially what the AI starts to optimize once it’s learned how to build a stable stack,” he said.
Multipick gripper optimizes motions
For smaller boxes, the AmbiStack robot is equipped with a vacuum gripper to pick up multiple boxes from a conveyor and place them on the stack in a single motion.
The gripper has multiple vacuum zones that can be selectively triggered to hold or release the boxes. This can improve throughput and enable the robot to handle a higher volume on the infeed conveyor, asserted Ambi.
The company plans to continue innovating gripper design and multi-zone actuation to enable the future picking of items such as pouches and soft-sided envelopes.
In contrast with traditional automation, which Ambi Robotics said is often expensive and inflexible, AmbiStack is configurable. It can be applied across multiple use cases such as building inbound pallets from floor-loaded trucks, assembling outbound pallets for retail distribution, and even stacking walled containers for parcel transport.
AmbiStack can also palletize during truck unloading
Ambi Robotics said it also sees applications for AmbiStack in truck unloading. The robot could be located near the loading dock and palletize boxes into unit loads as they come out of a floor-loaded trailer.
“A big part of the challenge there [with truck unloading] is that a lot of these operations don’t know what type of case they’re getting until they open that dock door for the first time,” said Mahler. “The ability to take that stream quickly in unstructured fashion and build pallets out of it is really of interest to a lot of these logistics companies.”
AmbiStack is entering limited production in 2025, with a select number of units available to early customers. Ambi Robotics said it expects units to sell out fast, with more than 10 customers waiting to get the system.
The company has begun securing orders for 2026, which it said demonstrates the strong market appetite for this technology.
“Logistics companies are under continuous pressure to deliver items faster and for lower cost, which puts a lot of strain on operations to maintain reliability while rapidly adapting for the future,” said Jim Liefer, CEO of Ambi Robotics. “AmbiStack can be rapidly configured to automate the repetitive motions involved in a huge variety of stacking and palletizing applications, driving greater accuracy and efficiency, while future-proofing warehouse operations.”