NVIDIA’s 14B DreamZero model passes 30 simulated trials but fails can-to-mug tasks
wandb · x · 2026-08-04
- W&B published a report evaluating NVIDIA’s 14B world-action model DreamZero on DROID tasks in NVIDIA Isaac Lab.
- The evaluation tested three manipulation/navigation tasks without changing the pretrained weights.
- Baseline results were uneven: cube → bowl and banana → container were 5/5, while can → mug was 0/5.
- In 30 trials with the cube shifted by 0–10 cm, the system succeeded every time.
- W&B notes this is a simulation-only external evaluation, not evidence of real-world robot performance.
More from Embodied
- Robot product builders should let users test on their own Hugging Face data — DominiqueCAPaul · 2026-08-04
- Google DeepMind releases Gemini Robotics 2 to push generality, whole-body control and dexterity — bousmalis · 2026-08-04
- Cartwheel ships Comic 4.2, a flagship perception model with better feet estimation — andrew_n_carr · 2026-08-04
- Concept mockup imagines a self-driving minivan with fold-out beds — LukeW · 2026-08-04
- ARPL makes llama.cpp adapt to ARM ISA and core topology at runtime — OpeningTough145 · 2026-08-04
- Multimodal spatial intelligence and physical AI pitched as the next frontier — richdotca · 2026-08-04