This work introduces InsertionWM, a model-based framework for robotic insertion. It combines wrist-mounted camera images with proprioceptive information in a single world model, training one generalist policy across many assembly tasks instead of separate specialist policies. On up to 90 geometrically diverse insertion tasks, the model achieved 56% zero-shot success on unseen objects with unknown geometry, compared to 7% for a model-free baseline. Performance improved as more training objects were added, and finetuning the generalist on held-out objects was more data-efficient than training from scratch, sometimes reaching better final performance. This suggests model-based RL is a more scalable route to generalizable assembly than specialist, model-free policies.
A single world model enables robots to insert unseen parts zero-shot
Training on 90 insertion tasks yields 56% zero-shot success, far outperforming model-free baselines.
Big Tech
Nicklas Hansen · Iretiayo Akinola · Yijie Guo · Jie Xu · Bingjie Tang · Hao Su · +4 more
NVIDIA · University of California San Diego · University of Southern California
Research Digest··2 min read
Thread:Robot World Models
In one line
A visual world model trained on up to 90 insertion tasks zero-shot assembles unseen objects with 56% success, far above a 7% model-free baseline.
Before you read this
Two ideas this paper builds onlatent world modelA learned internal, compressed representation of environment dynamics, enabling future-aware planning, generation, and reasoning directly in latent space.51 papers →reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.296 papers →
Read it as
Hansen et al. present InsertionWM, a framework that trains a single world model on up to 90 geometrically diverse insertion tasks using wrist-mounted camera images and robot proprioception.
The authors developed InsertionWM, a model-based approach that learns a visual world model from raw camera observations and proprioceptive data.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.