Positronic Robotics: smart cleaning with AIPositronic Robotics: smart cleaning with AI

In most hotels, bathrooms account for over 40% of cleaning time. Consider that’s usually the least appealing task for housekeeping staff and you have the perfect starting point for Positronic Robotics’ ambitious journey toward fully automated cleaning.

Bathrooms may not be glamorous, but they’re perfect testing grounds. Each one is relatively isolated, allowing the robot to work on one task while a human cleaner focuses on the rest of the room. Like eating an elephant one bite at a time, robots start with cleaning toilets, then move on to washing basins, mirrors, shower areas and so on. By beginning in places with consistent cleaning routines and high labor costs — like hotels — Positronic Robotics will refine its technology before moving on to commercial cleaning and, ultimately, household applications.

With a smarter approach to humans teaching AI in a controlled setting, slowly increasing the complexity of tasks over time, each new property added to the system accelerates training and broadens its scope. This collaborative model is a win-win for everyone: cleaners offload their least appealing tasks while robots learn directly from them in real hotel environments

Approaches and technologiesApproaches and technologies

Engineers at Positronic are building a fully end-to-end, Gen 3 robotics system based on machine learning, focusing on an imitation learning approach that trains robots to perform tasks by picking up from human demonstrations.

For reference:

• Gen 1 (Open Loop): Follows pre-set commands without environment feedback, often operating in fenced areas for safety.

• Gen 2 (Closed Loop): Uses sensors and mapping to adjust actions in real time, but relies heavily on predefined algorithms.

• Gen 3 (AI-driven): Employs neural networks to process sensor data directly and adapt quickly to changing conditions.

How does it work?How does it work?

To train the robot, people demonstrate how to perform tasks — like wiping a sink — over and over, and the model learns by observing these actions. As it sees more examples, the robot becomes capable of cleaning a range of objects in diverse environments with different kinds of contaminants. A high-level planning system, powered via modern VLM like GPT-4o, then decides how to switch between tasks and tools.

At the core of the robot’s control system is the Action Chunking Transformer (ACT), which merges features from a visual encoder and robot states with a transformer encoder-decoder model to predict the next steps. Beyond simply anticipating the next position, ACT predicts an entire Action Chunk, which helps mitigate the compound error effect. This approach is inspired by psychological theories on action chunking, which describe how sequences of actions are grouped together to improve efficiency. By structuring movements in this way, the robot can execute tasks more smoothly and with greater precision. Additionally, ACT is trained as a Conditional Variational Autoencoder (CVAE) to capture the variability in human data, such as pauses during demonstrations.

Unlike in the original ACT paper, instead of relying directly on robot joint states, Positronic represents automated movements in a Cartesian space, which allows to use the same policy for different robotic arms. This approach improves flexibility, ensures consistency across various hardware setups and allows demonstrations to be decoupled from specific hardware configurations. For modeling rotation, engineers use a specialized method that enhances stability and accuracy, avoiding the limitations of traditional quaternion-based representations.

In the midst of development, Positronic tackles all sorts of challenges. Increasing movement speed is a big priority, so that robots perform cleaning tasks at least as quickly as humans. The team also developed specialized cleaning attachments that both machines and people can use, making it easier to teach the robot. To mention another important achievement, Positronic recently trained the robot to maintain full surface contact while applying the necessary pressure to not just to glide over surfaces but to scrub off stubborn stains.

Using Nebius AI Cloud for model trainingUsing Nebius AI Cloud for model training

Computing: Positronic uses several virtual machines hosted on Nebius’ infrastructure through the Compute Cloud service to train its models. The startup relies on NVIDIA H100 GPUs for model training. Each training session takes place in a containerized environment on Nebius servers and is optimized through distributed parallel computing. Data is transferred to the servers and runs in containers with minimal manual configuration. A single server is dedicated to one training cycle, which lasts about 12 to 16 hours.

Data organization: The storage and management of training data also utilizes Nebius’ services. Currently, researchers use around one terabyte of data for ongoing training, but this volume is expected to grow significantly. During training sessions, data is collected via VR tools and sensors, then processed and prepared for use in the models.

Flexibility and automation: Training processes are automated with scripts that manage the start and end of each task. After training is complete, the server automatically shuts down to conserve resources. The company is exploring the possibility of running simulators in a virtual environment on Nebius servers for parallel model training in the future.

Plans for the futurePlans for the future

Right now, engineers at Positronic Robotics are focused on scaling up the data collection process. To achieve this goal, they plan on equipping cleaning staff with specialized tools, enabling the model to learn from a wide array of real-life examples and scenarios.

Another important next step will be the integration of simulators to accelerate model validation and virtual-environment training. This will allow multiple robots to be trained in parallel, improving their performance metrics and testing the resilience of their solutions under various conditions.

Looking ahead, Positronic is also exploring the possibility of running these simulators on Nebius servers, creating a virtual environment for parallel model training at a much larger scale.