Robotics & manufacturingResearch findings
One Policy for 100 Household Tasks
Original published 2026.11.11 03:06 KSTAdded to Blackdeck 2026.11.12 03:16 KST
What has changed?
A single VLA policy handles all 100 long-horizon household tasks efficiently.
Key points
- A unified vision-language-action policy solves 100 long-horizon household tasks in the OmniGibson simulator.
- Extended training to 200k steps with semantic stage warm-starting and noise reduction improves task performance.
- On 14 tasks, a GPT-6 Astra agent achieves a Q-value of 0.639, significantly outperforming the base policy.
Connect it to your work
The model was trained for 200k steps with semantic stage warm-starting and reduced gradient noise.
What remains unproven
The abstract reports simulation results; real-world performance is not established by this input.
Originals used for this note
Links open the original publisher in a new tab.