Aether AI

Aether AI is a San Diego, California company developing causal world models: AI models intended to represent cause-and-effect mechanisms, reason about interventions and counterfactuals, and predict the consequences of actions, rather than relying only on statistical correlations in training data. The company names physical AI and robotics as the first application domain for this technology, with scientific discovery (biology, medicine and longevity research) as a longer-term area.

The company was founded by Prof. Biwei Huang, an Assistant Professor at the University of California San Diego whose research focuses on causal discovery and machine learning. Huang is associated with the open-source causal analysis tools Causal-Learn and Causal-Copilot. Aether AI states that its founding team works in causal discovery, causal AI, causal foundation models, causal reinforcement learning, agentic systems and foundation model training. The contact address published on the company website is 8910 University Center Lane, Suite 400, San Diego, CA 92122.

In June 2026 Aether AI announced a USD 20 million seed round led by MPCi, with participation from Inno Angel Fund, SWC Global, Unity Ventures and other institutions. The capital is to be used for research and development of causal world models, expansion of the engineering and scientific team, and initial commercial deployments in physical AI and robotics.

The company publishes its research results on its blog. In September 2026 it introduced CausalWM, which it describes as its first embodied causal world model. CausalWM uses causal chain-of-thought reasoning to predict motion and geometry before generating future video frames; the company reports a score of 66.04 on TriWorldBench (ranked first) and a score of 89.9 in the PAI-Bench robot domain, stating that it outperforms Cosmos 3 Super. Other published work includes SCAR (self-supervised continuous action representation learning, providing a unified latent action interface across robot embodiments, August 2026), CD-LAM (causal debiasing for world models, reported to reduce action-following error by more than 30% with ten times less post-training, July 2026), a study on learning object manipulation through contact geometry in which an air-hockey robot improved from 5/20 to 12/20 successful trials, task-centric world models, and RSIAgent, an agent for recursive self-improvement evaluated on OSWorld 2.0 and Agents' Last Exam. These results are self-reported and have not been peer reviewed.

As of September 2026 Aether AI had announced no customers, priced commercial products or named robotics partners; it is a robot foundation-model and world-model developer at the research-to-early-commercialization stage.