Insider Brief
- Flexion Robotics introduced Reflect v1.0, a robotics intelligence platform designed to help humanoid robots complete longer, multi-step tasks without human operation during execution.
- The company demonstrated the platform through a workplace delivery task in which a humanoid robot retrieved a snack parcel, used stairs and an elevator, unpacked the box and placed items into a drawer.
- Flexion said Reflect v1.0 combines mission control, motion planning, whole-body control and runtime software, with reinforcement learning raising completion in an internal 16-step mission evaluation from 38% to 90%.
Swiss startup Flexion Robotics has introduced Reflect v1.0, a robotics intelligence platform it says is designed to help humanoid robots complete longer, multi-step jobs without human operation.
According to the company, the platform was demonstrated through a workplace delivery task where a humanoid robot was given the instruction, “A parcel with snacks has been delivered for Flexion. Retrieve it using the stairs and come up using the elevator. Then unpack it and place the items into the empty drawer on the shelf in the snack area.”
Flexion said the robot completed the mission autonomously, including finding the package, moving through a multi-floor building, handling doors and elevators, using tools to open the box and adjusting when parts of the task changed.
Flexion pointed out one barrier in robotics is reliability over time. Robots can often perform individual actions in short demonstrations. The harder problem is linking many actions together in a real environment, where a small navigation error, missed grasp or mistaken scene reading can cause the full mission to fail.
Reflect v1.0 is Flexion’s attempt to connect those pieces into one system as the platform combines mission-level reasoning, perception, robot motion, whole-body control and runtime software that manages communication, logging and safety checks.
How Does it Work?
Reflect v1.0 is built around four main layers:
- Mission control: A custom vision-language model sits at the top of the system, taking the user’s instruction, watching the robot’s camera view and deciding what the robot should do next as the task changes.
- Motion layer: Beneath it, a vision-language-action model trained on real-world data works with reinforcement-learning skills to turn what the robot sees into navigation, object handling and scene-interaction commands.
- Whole-body control: A real-time controller keeps the robot balanced and coordinated while it walks, carries objects or uses its arms.
- Runtime system: Supporting software manages communication, low-latency inference, process isolation, logging and safety checks.
Flexion said reinforcement learning was critical to making the mission controller more reliable. In an internal 16-step mission evaluation, supervised fine-tuning alone produced a 38% end-to-end completion rate, while adding reinforcement learning raised completion to 90%.
Below the mission controller, Reflect uses a motion layer that turns decisions into physical actions such as opening doors, calling elevators, navigating around obstacles and picking up objects. The company said the robot’s box-handling policy could pick up boxes ranging from 100 grams to 3.5 kilograms and adjust when objects were moved or initially missed.
The system also includes whole-body control software, called Reflex, that helps the robot maintain balance while walking, carrying items or interacting with objects. Flexion said the controller has supported more than 100 consecutive stair traversals and can help the robot remain stable under disturbances.
Flexion also built its own communications and runtime system, called FlexComm, to support the platform. The company said FlexComm delivered up to 40% faster same-host communication than ROS DDS and a 30% CPU improvement compared with ROS2 DDS, while improving resilience during network disruptions.
Limitations
The company said Reflect v1.0 still has limits. The system operates within a bounded set of tasks, some objects remain difficult to grasp, the mission controller can still make wrong assumptions from visual input and recovery behaviors do not cover every failure mode.
Still, Flexion said the release marks a step toward humanoid robots that can complete practical work over longer periods.
“The presented mission is an important milestone for us because these components are no longer evaluated only as isolated clips,” the company wrote in its announcement. “They can now be composed into a system that acts through time, interacts with the world, changes its plan, and continues to make progress under real constraints.”
Image credit: Flexion Robotics
