Robostral Navigate, an 8-billion-parameter model built entirely in-house, uses a single RGB camera to follow plain-language instructions through complex indoor environments. On the R2R-CE validation unseen benchmark, it achieved a 76.6% success rate, beating the best multi-sensor system by 4.5 percentage points while relying on cheaper, simpler hardware.
The model processes RGB images and a text instruction-such as "Leave the lobby, walk through the corridor, enter the supply room, and stop to face the second shelf"-and moves the robot through the space without any depth sensors or LiDAR. It outperforms the best single-camera approach by 9.7 points and the best system using depth or multiple cameras by 4.5 points.
How pointing-based navigation works
Robostral Navigate predicts the next movement by pointing to image coordinates in the robot's current camera view, along with the desired orientation. Because the model points directly to image coordinates, it stays accurate even when camera intrinsics or world scale vary. When the target is outside the field of view, the model falls back to metric displacement commands, such as "Move 2 meters forward, 1.5 meters to the left, and turn 25 degrees left."
The model was built on a vision-language model specialized for grounding tasks like pointing, counting, and object localization. Navigation emerged as a natural extension: once the model understands where things are, it learns how to move.
Training efficiency with prefix-caching
The team built a data generation pipeline entirely in simulation, producing approximately 400,000 trajectories across 6,000 scenes. A key training innovation is prefix-caching, which uses a tree-based attention-masking strategy to compress an entire episode into a single sequence. This allows training on all time steps in one forward pass without information leakage, reducing the number of training tokens by 22×. In practice, training runs that would take months complete in days.
This technology enables applications across manufacturing, delivery, logistics, and hospitality, making it a notable advance in AI Agents & Automation.
Online reinforcement learning improves recovery
After supervised training, the model's performance was further improved using CISPO, an online reinforcement learning algorithm. This lets the model learn from trial and error, recover from failures, and acquire exploratory behaviors, mitigating the distribution shift problem of vanilla behavior cloning. The RL stage alone added 3.2% to the success rate, and the team reports no plateauing, suggesting further gains with more training.
Works across robot types and camera setups
Robostral Navigate runs on wheeled, legged, and flying robots, and generalizes across different robot sizes. It also handles variations in camera intrinsics without issue. The model handled real-world obstacles and moving people in a working office during fully autonomous long-horizon tests, despite never seeing those specific environments during training.
Toward a unified embodied agent
The release marks a first step toward a general-purpose embodied AI. The team is actively hiring research scientists and engineers to expand the robotics effort, aiming to bring navigation to robots everywhere-in offices, homes, commercial buildings, and outdoor spaces.
Why this matters for General, IT and Development
For IT and development teams, the model's training efficiency-reducing token counts by 22×-exemplifies the kind of practical gains that matter in AI for IT & Development. The compact 8B size and reliance on a single RGB camera lower the hardware and compute barriers for deploying autonomous navigation. Combined with its ability to generalize across robot types and environments without retraining, Robostral Navigate offers a practical path to adding embodied AI capabilities to robotics projects.
Your membership also unlocks:
