Today we’re introducing Robostral Navigate, the first model built for embodied navigation. This is an 8B model that accepts RGB images and plain language commands to move the robot within its environment.
“Exit the lobby, walk down the hallway, enter the supply room, and stop at the second shelf.”
Other models often employ depth sensors, LiDAR, or multiple cameras working together to accomplish such tasks. Although Robostral Navigate uses only one regular RGB camera and no depth sensor, it invisibly achieves 76.6% in R2R-CE (Room-to-Room in Continuous Environment) validation, a benchmark for following instructions in non-training environments. The result is that it outperforms the best single-camera approach by 9.7 points and outperforms the best system using depth or multiple cameras by 4.5 points, despite using neither.
navigation
Our model is designed for robot navigation, allowing robots to autonomously navigate complex environments such as offices, residential and commercial buildings, and outdoor environments.
Robostral Navigate runs completely autonomously with one long-term directed route through the working office.
This technology enables numerous applications across manufacturing, distribution, logistics, and hospitality, making it one of the most sought-after capabilities for customers today. If you give Robostral Navigate one instruction, the robot can complete the entire task automatically, navigate live spaces full of previously unseen people and obstacles, and adapt to any setting.
highlights
-
R2R-CE cutting-edge performance
-
It does not use LiDAR or depth sensors and works with a single RGB camera.
-
8B model, built in-house and fully trained in simulation
-
It runs on wheeled, legged, and flying robots and is generalized across robot sizes.
-
Strong against camera-specific differences
-
Token-efficient training with prefix caching
Navigation by pointing
Given a task and observation history, Robostral Navigate predicts where the robot should move next. pointing: Estimate the image coordinates of the target location within the robot’s current camera view and the desired orientation upon arrival. Unlike commands that rely on metric displacement, pointing makes the policy naturally robust to changes in camera nature and world scale.
However, this method cannot accommodate cases where the target location is outside the current field of view. If no pointing is applied, the model reverts to displacements in the robot’s local coordinate frame as follows:
“Move forward 2 meters, left 1.5 meters, and rotate 25 degrees to the left.”
Built from scratch
Robostral Navigate is built entirely in-house and does not rely on any existing open source VLM.
The model is initialized from a vision language model specialized for grounding tasks such as pointing, counting, and object localization. Navigation emerges as a natural extension of these capabilities. Once they understand where things are, they learn how to move them.
We built an efficient data generation pipeline entirely in simulation. This allows for rapid iteration of the data, resulting in approx. 400,000 trajectories collected across 6,000 scenes.
Efficient supervised training
A key element of Robostral Navigate is an efficient training algorithm based on prefix caching. Our method uses a tree-based attention masking strategy to compress the entire episode into one sequence, allowing training for all time steps in a single forward pass while preventing information leakage between time steps.
Compared to training with one sample per time step, our approach reduces the number of training tokens by: 22× while maintaining all learning signals. In practice, this method Convert training runs that take months into runs that complete in days.
online reinforcement learning
Leverage knowledge from large-scale post-training LLMs using online reinforcement learning to improve Robostral Navigate performance. After the supervised training phase, we further improve the model’s performance using CISPO, an online reinforcement learning algorithm. This allows the model to learn from trial and error, recover from failures, and acquire exploratory behavior, effectively mitigating the problem of distribution shifts caused by cloning vanilla behavior. This alone increased the success rate by 3.2%. We don’t see a plateau, so we’re confident more training and more experimentation will continue to push this number up.
what’s next
Robostral Navigate is just the first step toward integrated bodily agents.
We believe that navigation is a fundamental function of general purpose robotics. Robostral Navigate demonstrates that by combining large-scale simulation, efficient training, and strong grounded prior learning, state-of-the-art embodied navigation can be achieved with a compact model and a single RGB camera.
Start your journey towards bringing frontier AI to life and talk to our team.
By the way, we are hiring!
While the release of our navigation model represents an important step forward, our journey is far from over. Our goal is to enable robots to autonomously navigate complex environments such as offices, homes, commercial buildings, and outdoor spaces, but there is still much work to be done. We are actively expanding our robotics team and are looking for talented research scientists and engineers who share our aspirations.
If you are interested in joining our mission to provide seamless navigation for robots wherever they are, we welcome you to apply to join our team.
By Théo Cachet, Arjun Majumdar, Srijan Mishra, Thomas Chabal, Chris Bamford, Elliot Chane-Sane, Benjamin Tibi, Ludovic Ho Fuh, Olivier Duchenne – AI Science Robotics
