Zhejiang University Releases RoboSPA Benchmark to Stress-Test VLA Models on Long-Horizon Tasks
zju · hf · 2026-09-09
Zhejiang University released RoboSPA on Hugging Face, a large-scale robotic manipulation benchmark evaluating vision-language-action models beyond simple scenes and short-horizon tasks.
- Focuses on fine-grained spatial reasoning and long-horizon procedural planning
- Uses progressively harder task variants to expose generalization gaps
- Responds to concerns that current VLA models only excel in simple, short-horizon settings
More from Embodied
- Adding Greek to a Cosmos3 VLA policy: bilingual training helps but lags far behind English — KIEFERSA · 2026-09-09
- OpenAI likely to make a big robotics move, Darpinian predicts ahead of EoY — Darpinian · 2026-09-09
- At ECCV 2026, researcher says ask 'what is an action,' not 'what is a world model' — dimadamen · 2026-09-09
- SoftBank backs Gravis Robotics with $200M Series A at >$1B valuation for excavator autonomy — lukas_m_ziegler · 2026-09-09
- Apple fall event preview: iPhone Duo foldable and biggest product pipeline in years — APPSO · 2026-09-09
- BYD flash-charges 20% to 97% in 12 minutes at -30°C, with 10,000+ stations in 332 cities — chrisfirst · 2026-09-09