Most VLAs lose their base VLM's vision-language skills during robot pre-training, eval shows
DJiafei · x · 2026-11-16
By isolating the VLM weights inside open-source VLAs and benchmarking them on 8 vision-language tasks, an analysis shows most VLAs dramatically degrade their base model's skills after robot pre-training: OpenVA drops 50.3 pts vs Prismatic 7B, NVIDIA GR00T N1.7 drops 65.3 pts vs Cosmos-Reason2, and G0.5 drops 33.5 pts vs Qwen3.5-2B-Base. Exceptions exist: π0-5 gains +4.1 pts over PaliGemma 3B, and MolmoAct2 loses only 1.5 pts. Models fall into three categories: near-total loss, partial retention, and near-full retention—highlighting a systemic capability-erosion problem in the VLA pipeline.
More from Embodied
- A beginner's guide to training VLAs: microVLA dev shares low-compute training intuitions — joemeno · 2026-11-16
- NVIDIA AI Day Seoul Draws 2,000+ Attendees Focused on Physical AI and Agents — NVIDIA Blog · 2026-11-16
- Ring Launches $1,000 Ring Rover Cam, an Autonomous Patrol Security Robot — emmanuelvivier · 2026-11-16
- Amazon to Equip 20,000+ Drivers With Smart Delivery Glasses by End of 2027 — emmanuelvivier · 2026-11-16
- Tesla Robotaxi geofences now run fully autonomously 24/7, Cybercab consumer sales teased — JOBhakdi · 2026-11-16
- Tesla may ship a steering-wheel-free car by end of year, with huge transit implications — mobileraj · 2026-11-16