Three-Stage Training Aligns Vision, Language, And Robot Actions
Reactions from ranked influencers
3 posts
馃У 5. Block-wise decoding generates several weakly connected action tokens together instead of producing all 32 tokens one by one. On 1 NVIDIA RTX 5090 GPU, Astribot reports latency falling from 253.66ms to 93.53ms, a 2.71x speedup.
馃У 7. Lumo-2 jumps far beyond Lumo-1 across many embodied reasoning tasks while staying competitive with models built mainly for vision-language benchmarks. The robot training may actually have made its grasp of space and physical scenes more useful.
Technical Report: https://www.astribot.com/en/research/Lumo2 Project page: http://www.astribot.com/research/Lumo2 馃搫 Paper: https://arxiv.org/abs/2607.11270