
SE3 Labs turns real factory and hotel workflows into repeatable robot tasks for model teams to train and evaluate policies.
SE3 Labs' real-world benchmarks and failure analysis could raise growth because each evaluation compares checkpoints on same tasks, measures improvements and turns remaining failures into targeted data for next cycle, prompting model teams to repeat evaluations; wider use confirms this; fewer weaken. RoboDojo Real builds ten RoboDojo simulation tasks in the real world, two per generalization, memory, precision, long-horizon and instruction-conditioned manipulation, on a bimanual station scored by milestone.
In a 20-episode comparison, the code policy completed 6 placements, while agentic control kept a lift in 18 of 20 episodes and completed the placement in each. 61 episodes hit at least one milestone while 47 stalled until the timeout. The company plans to set up physical sandboxes at partner sites so policies can train on the exact tasks and conditions they will face in deployment.
Value in physical AI may increasingly shift from hardware, which currently attracts most investment, to software, deployment, and field operations, favoring providers in those later stages over hardware suppliers.
The richest physical AI insights come from field deployment, where systems encounter daily life's uncertainties and messiness. Simulation data lacks the robustness of physical-world insights, and each trajectory a robot learns from requires physical performance, capture, and careful structuring before it can teach anything.
Physical capacity is the binding constraint on SE3 Labs' growth: each real-world attempt ties up a robot and a task setup, requires a dependable success check, and needs a reset before another trial. Mark, Conan and Paige are the team behind SE3 Labs, and because real-world evaluation remains costly and difficult to scale, the cost of scaling evaluation is the main execution risk.
SE3 Labs runs policy deployment across standardized robot workcells in controlled labs, factories, and other operating environments. A trained operator completes the task and returns control after a policy fails or reaches a safety boundary. The company plans physical sandboxes at partner sites where policies can train on the exact tasks and conditions they will face in deployment.
SE3 Labs offers robot-model teams a real-world evaluation loop instead of self-built infrastructure. Teams supply a model checkpoint or inference endpoint connection and define the task, test environment, and success criteria, while SE3 deploys policies across standardized robot workcells in controlled lab, factory, and other operating settings, managing hardware, sensors, task configuration, safety limits, scheduling, and run execution under one protocol applied across test stations and policy versions for consistent result comparison.
When a policy fails or triggers a safety boundary, a trained operator completes the task and returns control, and SE3 retains the full handover as an unbroken correction trajectory. Raw recordings become synchronized, quality-controlled datasets segmented by task and outcome and returned in one consistent format. RoboDojo Real offers physical builds of ten RoboDojo simulation tasks—two apiece across generalization, memory, precision, long-horizon, and instruction-conditioned manipulation—executed on a bimanual station and scored by milestone.