What we are building and what we are learning: notes on in-context robot learning, write-ups of the work, and news from the team.
4 min read
Vector Level 1 concluded with an 87.9% success rate across 16 two-arm robot tasks in unseen scenes, each prompted by a single demonstration, compared with 87.2% for fine-tuned π0.5. Across 68 duels, miners competed by submitting model weights under matched evaluation conditions. Level 2 is now testing robustness to changes in lighting, textures and table height.
7 min read
WorldGen turns a text prompt into a traversable, editable 3D environment by combining scene planning, procedural structure, generative appearance, and 3D reconstruction. Unlike pixel-only world models, it outputs explicit geometry and objects that can be used inside real simulation pipelines. Its real value is making world creation scalable enough for robotics training and deployment.
8 min read
The next frontier of robotics AI is turning intelligence learned in simulation into reliable behavior in the physical world. Sim2Real is becoming a core stack—combining domain randomization, system identification, digital twins, learned dynamics, and real-world feedback to bridge that gap.
5 min read
Behavior Prompting Policy (BPP) lets robots learn new tasks from a single demonstration at inference time, using the demo as context rather than retraining the model. Its strongest results suggest that behavioral diversity matters more than repeated demonstrations, enabling robots to recombine familiar skills into new tasks without gradient updates.
6 min read
Zero-WAM brings in-context learning to robotics by using human demonstration videos as task prompts, achieving strong zero-shot cross-task results with a large world-action model and HumanGen synthetic data. An important step toward robotic ICL, but current evidence remains domain-specific, computationally heavy, and short of universal zero-shot robot learning.
11 min read
Robotics is evolving from task-specific imitation toward generalist models that can adapt, learn from experience, and improve after deployment. The next robotics frontier isn’t just better pretrained models—it’s robots that can attempt, observe, remember, practice, and improve from their own real-world experience.
7 min read
Robot benchmarks have become much better at varying environments and scenes, but they still often rely on predefined task sets. The next step is to generate new, valid tasks at evaluation time, creating a stronger test of whether robots can truly generalize beyond benchmark-specific patterns.
7 min read
A policy that meets an unfamiliar task ought to get better at it while it is working, not after a retraining run. This is what that means precisely, where the idea comes from, why the competition runs two tracks rather than one, and what a strong result would and would not prove.