Fine-tuned World Action Model Dream-1 Jumps 159%, Beats Frontier LLMs on Robot Tasks
paigeinsf · x · 2026-11-19
SE3 Labs fine-tuned the Dream-1 world action model on 500 real-robot demos across 10 tasks, lifting its score 159% from 8.5 to 22 and beating frontier LLMs on repeatable physical routines in a fifth of the time.
Key results:
- Box packing (toy car, shoe, toothbrush, hammer): Dream-1 scored 54.5 vs π0.5's 51 and Fable's 39; Dream-1 packed all four items in 6 of 10 runs while Fable managed it once and rarely finished within its 15-minute limit (5x Dream-1's time budget).
- Stacking toy: Dream-1 35 > Fable 31 > π0.5 15.
- Randomized demonstrations spread skills across the whole table: edge reach went from 4 of 16 toy peppers to 11 of 14, and it even learned an ignored rule — lining up cubes using only a ruler — going from a zero-shot score of 1 to 41 after fine-tuning.
Takeaway: LLMs know what to do, WAMs know how to do it; the best robot brain will likely combine both.
More from Embodied
- SE3 Labs Releases A Wider Dream Eval, Opens 500-Episode Robot Demo Dataset on Hugging Face — paigeinsf · 2026-11-19
- LLMs zero-shot beat fine-tuned world action models on 7 of 10 robot tasks — paigeinsf · 2026-11-19
- Memory and Reasoning Remain LLM Territory: Fable Beats Dream-1 64 to 9 on Cup Task — paigeinsf · 2026-11-19
- The Secret Wasn't a Bigger Model: 500 Randomized Demos Lift Dream-1 Score 159% — paigeinsf · 2026-11-19
- Fine-tuned Dream-1 world action model jumps 159% on real-robot benchmark at 1/100th the cost — paigeinsf · 2026-11-19
- Robotics founder: "sim2real is BS" — labs buying data is a flash in the pan — Scobleizer · 2026-11-19