Intelligence should be verifiable, grounded, and able to act in the world.

Manifesto

Physical intelligence depends on more than a capable policy. It requires scalable data collection, representations that connect perception to action, and learning systems that transfer across tasks, environments, embodiments, and the simulation-to-reality gap.

Current direction

• Reading the Physical Intelligence π-series and GR00T to understand generalist vision-language-action policies, open-world generalization, and cross-embodiment transfer.
• Studying how teleoperation, human video, robot experience, and synthetic data can be collected and combined into scalable training mixtures.
• Connecting prior work in robotics, 3D perception, reconstruction, and geometry to sim-to-real transfer, adaptation from corrections, and failure diagnosis in long-horizon physical tasks.

Manifesto

A fluent rationale isn't trustworthy simply because it sounds coherent. A grounded system should be right about the answer and expose whether its reasoning depends on the correct evidence.

Current direction

• Submitted work through an evaluation competition; details and results go up when I can share them.
• Working down the post-training line: PPO, then GRPO, its DAPO variant, and PAPO — which adds an implicit perception loss on top of GRPO so a multimodal model has to use the image instead of reasoning around it.
• Exploring corruption augmentation, gated rewards, and grounded verification for failures caused by degraded or shifted visual evidence.

Manifesto

A single score hides the distinction between perception, generation, tool use, evaluation, and verification failures. An evaluation becomes useful when it identifies the behavior that broke and helps decide what to change next.

See the related system →

Current direction

• Designing pre-production quality gates with automated grading, adversarial probing, and red-teaming.
• Turning post-production failures into reproducible, component-level diagnoses instead of undifferentiated error logs.
• Testing whether synthetic questions resemble real user intent closely enough to expose meaningful weaknesses.