Embodied-R1.5: 8B-Param Model Tops 16 of 24 Embodied VLM Benchmarks

Embodied-R1.5 is an 8B-parameter Embodied Foundation Model trained on 15B+ tokens across cognition, planning, correction, and pointing tasks using a multi-task balanced RL recipe. A Planner-Grounder-Corrector closed-loop framework enables autonomous self-correction on long-horizon physical tasks, achieving SOTA on 16 of 24 embodied VLM benchmarks.