Simulation Training and Efficiency Boosting Methods

MistralAI · x · 2026-07-08

Officials state that Robostral Navigate is trained entirely in simulation using about 400,000 trajectories across 6,000 scenes. The post also highlights a prefix-caching recipe that reduces training tokens by a factor of 22, compressing a process originally measured in months down to days. Additionally, the online reinforcement learning method CISPO further improves the success rate.

Related event: Mistral Unveils Robostral, an 8B Embodied Navigation Model(12 posts)→

Original post →

More from Embodied

Embodied channel →