T2-Z adds a learnable environment latent to a robotic world model to capture hidden factors such as friction, mass or task rules, without labels for what those factors are. Training alternates between learning shared action-conditioned dynamics with the latents fixed and optimising the latents with the model frozen; at test time a fresh latent is fitted from one or a few interactions. Across 7 simulated settings and 4 real tasks, object prediction error falls from 15.20 to 7.00 and action success rises from 33.73% to 72.96%.
Relevance
Adapting from a single interaction is what a robot needs when the field differs from the lab. Read for long-term autonomy and foundation models for robotics; how the latent behaves when the environment keeps drifting is open.
Links
- arXiv
- Subfield: Robot learning, Semantic mapping