The publication record

Physical AI is not one discipline. A system that works on a factory floor needs perception, representation learning, planning and control to hold together at once — so our record spans all four. Everything below is peer-reviewed and public. Each entry links to the paper.

11 of 11 shown

Robotics & Manipulation

1 paper

Learning to grasp anything by playing with random toys.

Training a grasping policy on randomly assembled shape primitives generalises to unseen real objects, with no in-domain data collection.

Embodied AI

2 papers

Hyperbolic planning and curiosity for crowd navigation.

Planning in hyperbolic space gives an agent a better structural prior for navigating dense human crowds.

Following the human thread in social navigation.

Tracking human intent as a continuous signal, rather than as obstacles, produces socially legible robot motion.

Computer Vision

4 papers

Online mistake detection in procedural egocentric videos.

Detects a procedural error as it happens from a first-person view — the perception half of any assistive or inspection system.

Hyperbolic active learning for semantic segmentation under domain shift.

Chooses which pixels are worth labelling when the deployment domain differs from the training one — directly a data-cost problem.

Human motion unlearning.

Removes specific learned motions from a generative model without retraining it from scratch.

VIDEO UNLEARNING

Unlearning in video models via a low-rank refusal vector.

A low-rank edit suppresses unwanted generations, giving a practical control surface for deployed video models.

Multimodal & Foundation Models

4 papers

Compositional entailment learning for hyperbolic vision-language models.

Learns part-whole structure between images and text, so the model composes concepts instead of memorising pairs.

HYPERBOLIC MULTIMODAL

Hyperbolic learning with multimodal large language models.

Shows hyperbolic representations scale to multimodal LLMs, improving how hierarchy is encoded across modalities.

Selective resampling for expressive state space models.

Makes state space models more expressive at long range while keeping their linear-time advantage over attention.

Evaluating coding LLMs at one-million-token context windows.

A benchmark showing where long-context claims break down in practice on real codebases.

Where It Goes Next

Papers prove the method. Projects prove it runs.

WIZARD builds on the one-shot imitation work with UC Berkeley. T-Rex builds on the tactile and manipulation line. Both operate on real hardware.

WIZARD ALL PROJECTS