Kinisi Robotics: How a Startup Is Redefining Humanoids with Real-World AI + AMRs

In this talk

1. 1:44

   Why Kinisi pairs arms with AMRs

   After spending a decade on legged robotics research, Pierce argues that many warehouse jobs happen on level floors, where legs add complexity without helping. Kinisi combines cobot arms with AMR bases, drawing on navigation technology Pierce helped build at Bear Robotics. That lets the new company focus on teaching robots to perform tasks such as moving parts and handling boxes.

2. 4:11

   Bringing AI performance into real deployments

   Pierce says lab results around 80 or 90 percent reliability fall short of customer needs, which can approach 99.9 percent. He compares the challenge to bringing SLAM from research into warehouses, a process that took years and a large engineering team. Kinisi is concentrating on a narrow set of tasks, including picking up a box, applying a label, and placing it on a conveyor.

3. 5:21

   Training with demonstrations and synthetic data

   Kinisi uses Bear’s API for navigation, then applies reinforcement learning to longer tasks and imitation learning from human-controlled examples. Pierce says newer methods can turn a single demonstration into many synthetic training examples. The team records robot and box footage, then varies warehouse settings and box appearance to broaden the training data; he says about an hour of demonstrations can be expanded this way.

4. 9:52

   Lessons from scaling Bear Robotics

   Pierce recalls that Bear’s restaurant robot worked in an early prototype, but busy restaurants quickly introduced obstacles such as rearranged tables and crowded groups. As deployments grow, customer environments and tasks differ, creating pressure to customize each installation or maintain many software variations. Bear ultimately chose to be a robotics company, with navigation as its core strength, rather than limiting itself to restaurants.

5. 8:40

   Building a small team around new research

   Kinisi had six people at the time of the conversation and was recruiting through robotics conferences rather than formal university partnerships. Pierce says the team persuaded three researchers in imitation learning to join, including recent PhD graduates and postdocs. He emphasizes how quickly the field moves, with new techniques changing what is possible with demonstrations.