Swiss startup Flexion Robotics has come up with yet another amazing innovation in the field of humanoid robotics. The goal of this innovation is to make humanoid robots way more autonomous. What does that mean? Well, that means operating a robot will now require little to no human intervention. And those long and multi-step tasks will be performed by these robots autonomously.

Flexion Robotics has recently introduced a new robotics intelligence platform. And they’ve named it Reflect v1.0. The platform is designed to help robots perform complex and multi-step tasks with as little human intervention as possible. Wondering how it functions? So basically, this platform uses advanced artificial intelligence, reinforcement learning, perception, and whole-body control so that robots become capable of performing complex real-world missions from the very beginning till the end on their own.

To demonstrate how Reflect v1.0 works, Flexion gave a simple natural language instruction to this new system. The robot was assigned the task of receiving a snack parcel that was just delivered to the building. To reach that parcel, the robot was asked to use stairs, and in order to come back, the robot was asked to use an elevator. Not only this, but the robot was also asked to unbox the parcel and organize the snacks in an empty drawer. And guess what? As per the company, the robot actually performed all these tasks autonomously. From finding the package and moving through a multi-floor building, handling doors and elevators, and using tools to open the box and organize snacks, the robot did it all. And throughout the task, the robot was seen adapting to changing conditions, too.

Designed for Long-Horizon Autonomous Tasks

Do you know what’s one of the biggest challenges for humanoid robots? Long-horizon autonomy it is! Yes, that’s true. And the best part about Flexion Robotics’ Reflect v1.0 is that it highlights this challenge. You must have seen many robots that perform individual tasks like simply walking or carrying objects, etc. But combining all these tasks into one uninterrupted mission, that too without any error, still seemed difficult. 

Now this is exactly where Reflect v1.0 makes the difference. While designing this new system, Flexion did keep in mind this challenge. This new system integrates mission planning, robot perception, motion generation, whole-body control, and runtime management into one unified platform. Rather than following fixed commands, this system allows the robot to observe its surroundings and take action promptly. And if needed, the system is capable of adjusting its plan as well.

Key Features of Reflect v1.0

At the heart of the platform is a custom vision-language model that acts as the robot’s mission controller. It interprets natural language instructions, analyzes live camera feeds, and determines the next action as the mission unfolds. The system also uses a language-grounded semantic map that allows the robot to understand buildings and locate objects or destinations through text-based queries.

Beneath the mission controller is a motion layer that converts decisions into physical actions. Powered by a vision-language-action model and reinforcement learning-based skills, it enables the robot to navigate hallways, climb stairs, open doors, operate elevators, manipulate tools, and handle objects in dynamic environments. Flexion said these policies were trained using both real-world data and simulation, allowing the robot to adapt when objects are misplaced or scenes differ from previous experiences.

The platform also includes Reflex. Reflex is basically Flexion’s whole-body control system. It helps the robot in many ways, such as maintaining balance while walking and carrying objects. Plus, it also assists the robot in interacting with the environment. The company said the controller has successfully completed more than 100 consecutive stair traversals while remaining stable during external disturbances and contact-rich tasks.

Reinforcement Learning Improves Reliability

A major advancement in Reflect v1.0 is the expanded use of reinforcement learning across the entire software stack. Previous versions mainly applied reinforcement learning to motion skills, but the latest release extends it to mission-level reasoning and decision-making. According to Flexion, supervised fine-tuning alone achieved a 38% end-to-end completion rate during an internal 16-step mission evaluation. After adding reinforcement learning, that success rate increased to 90%.

In order to support deployment, Flexion has developed FlexComm too. Wondering what’s that? So basically, FlexComm is a custom runtime and communication framework. What it does is keep data flowing smoothly across the robot’s system and manage process isolation and logging. Moreover, its job is to monitor safety during operation as well. The company claims the system delivers up to 40% faster same-host communication and a 30% CPU efficiency improvement compared with ROS2 DDS, while offering greater resilience during network interruptions.

Current Limitations and Future Outlook

Despite the progress, Flexion acknowledged that Reflect v1.0 is not yet a fully general-purpose autonomy system. For now, the platform only operates within a defined range of tasks. Grasping a few objects still seems difficult. Plus, the mission controller also makes mistakes at times while making assumptions on the basis of visual input. Recovery behaviors also do not cover every possible failure scenario.

Even with these limitations, Reflect v1.0 represents an important milestone in humanoid robotics. By combining high-level reasoning with reliable motion control and continuous replanning, the platform moves humanoid robots closer to performing practical, extended tasks in real-world environments without constant human supervision.