Lab Projects

The Reasoning Scaling Law: Why Scaling Logic is 1000x More Important Than Scaling Pixels

We show that the AI industry has reached diminishing returns in visual fidelity and should pivot toward reasoning capabilities. Our key finding: while visual quality plateaus early, reasoning capabilities exhibit a distinct "emergence" phase — increasing reasoning-specific training data by three orders of magnitude enables models to move from pattern-matching to genuine generalization in physical scenarios. We also introduce VBVR-Bench, a verifiable, rule-based evaluation framework for video reasoning that uses structured, multi-choice tasks and programmatic verification to avoid bias in AI-graded assessments.

This work is part of our mission at EmbodyX to bridge the spatiotemporal reasoning gap and move AI from generating pixels to performing actionable, physics-aware reasoning.

Human Cognition in Machines: A Unified Perspective of World Models

Our latest report (led by Timothy Rupprecht and our team at the PAIR Center) introduces a unified framework for World Models grounded in Cognitive Architecture Theory. By auditing current SOTAs across video, embodied, and epistemic domains, we provide a roadmap to bridge the gap between machine and human-like cognition.

This theoretical framework serves as a core pillar for our work at EmbodyX.

A Breakthrough in Embodied AI: CMU and EmbodyX Enable Robots to "Think Before They Act"

CMU researchers led by Jun Liu, in collaboration with EmbodyX, introduced RARRL (Resource-Aware Reasoning via Reinforcement Learning) — a framework that enables robots to dynamically decide when to perform high-level reasoning versus executing actions directly. The system adjusts reasoning intensity based on task difficulty, environmental urgency, and available computational resources. In ALFRED benchmarks, RARRL achieved over 60% reduction in overall reasoning time while maintaining or exceeding the performance of traditional always-reasoning methods.

A collaboration between CMU and EmbodyX on resource-aware reasoning for embodied AI.