CMRI's Lion-ER-35B-A3B: open embodied reasoning MoE lifts all 10 benchmarks at ~3B active params
量子位 · wechat · 2026-11-04
CMRI's Lion Mountain AI Lab released Lion-ER-35B-A3B, an open-source embodied reasoning VLM with a MoE architecture: 35B total parameters, 3B activated per token, weights on Hugging Face. Instead of blindly scaling robot data, the team built a capability-driven data engine: an 8-domain/39-skill capability blueprint, "capability compilation" into evidence-backed training questions, three supervised curation routes, and an eval-driven loop that turns model failures into targeted new data from a 4.3M-record pool (312K training samples). The model improves all 10 public benchmarks (65.29 avg score), beating its Qwen3.6 base by +9.30 on Where2Place and +12.12 on ERQA trajectory understanding, with gains on camera-motion, trajectory, cross-view spatial, and task-progress reasoning.
More from Embodied
- Baidu's Apollo Go wraps two-week public robotaxi test rides in Tung Chung, Hong Kong — Baidu_Inc · 2026-11-04
- imec neural probe integrates photonics directly into 130nm CMOS backend — jwt0625 · 2026-11-04
- Chris Paxton: interactive robot sim is the future but glaringly data-limited — Scobleizer · 2026-11-04
- DreamX: video diffusion models as universal latent action learners for robots — kaist-ai · 2026-11-04
- RoboTryOn turns human videos into robot demos, lifting success rate from 2% to 60% — kaist-ai · 2026-11-04
- INMO Future glasses: dual 1080p displays, 125g, $799 early-bird on Kickstarter Dec 10 — Scobleizer · 2026-11-04