Tsinghua Team's PrismWAM Tops RoboTwin 2.0, RoboDoctor Leads RoboDojo
On October 24, 2026, a research team led by Professor Li Yiming at Tsinghua University's School of AI, founder of Licing AI, unveiled two Physical AI research outcomes: the world action model PrismWAM and the agentic framework RoboDoctor.
PrismWAM explores how to represent physical states, predict state changes, and generate actions. It ranked first on the RoboTwin 2.0 benchmark with an average success rate of 83.67%, and achieved 75.5% on the Clean-to-Random high-difficulty evaluation. RoboDoctor, which focuses on diagnosing, correcting, and accumulating experience from execution feedback, scored 57.1 across all 54 tasks in RoboDojo, leading the current comparison.
Both works advance the scaling of Physical AI from the model and system levels. Without additional action-data pretraining, PrismWAM outperformed methods that rely on large-scale action-data pretraining, including OpenWAM, Xiaomi Robotics-0, and π0.5. RoboDoctor surpassed HexaShell-Preview (53.18) and WorldScape-R²SI (46.58), the top two on the official RoboDojo leaderboard, by 3.92 and 10.52 points respectively.
