Adding Greek to a Cosmos3 VLA policy: bilingual training helps but lags far behind English
KIEFERSA · hf · 2026-09-09
This HF post measures language transfer in robot vision-language-action policies by feeding machine-translated Greek instructions into a Cosmos3-based VLA model.
- The experiments expose measurement pitfalls in cross-lingual robot evaluation
- Bilingual training improves performance over monolingual baselines
- Greek performance still trails English by a wide margin, showing strong language bias in VLA models
Useful methodology and both positive and negative results for anyone building multilingual robot instruction following.
More from Embodied
- kornia-rs: Rust Low-Level 3D Vision Library Outperforms OpenCV and NVIDIA VPI — edgarriba · 2026-09-09
- ECCV 2026 workshop paper improves FoundationPose by enforcing a coherent scene — ducha_aiki · 2026-09-09
- Give the agent a physical lab and let it write the simulator: a fast path to robotics AGI — YouJiacheng · 2026-09-09
- Apple Watch Series 12 and Ultra 4 rumored to get privacy-focused Recaps AI that summarizes your day — Scobleizer · 2026-09-09
- Tsinghua's Z-Trans Bets Embodied AI's Next Battle Is Post-Deployment Self-Evolution, Not Data Scale — 新智元 · 2026-09-09
- OpenAI likely to make a big robotics move, Darpinian predicts ahead of EoY — Darpinian · 2026-09-09