Event

XPENG ROBOTICS Releases Embodied VLM Capek 0.5

1 reports · 1 sources

XPENG ROBOTICS has released the embodied vision-language model Capek 0.5, training separate experts for four categories of execution capabilities—spatial reasoning, temporal understanding, action guidance, and state verification—then integrating them into a single model using TIES merging and routed MOPD distillation. The model comes in two scales, 2B and 35B-A3B: 35B-A3B is an MoE that activates only about 3B per inference; 2B is a dense model. XPENG ROBOTICS states that on the strictly held-out EmbodiedBench (EB-HAB / EB-ALF, 300 multi-step episodes each) that was not used in training, the two scales improved on 30 and 28 of 34 benchmarks, respectively. These results come from benchmark evaluations self-reported by XPENG ROBOTICS, and the evaluation will be extended to longer-horizon interactive tasks and physical robot scenarios.