University of Southern California
Major research university in Los Angeles whose faculty are studying AI's effects on human language and communication.
“We present ViBe, a post-training framework for adapting motion trackers to perceptive control tasks.”
Source→“Colosseum V2: Benchmarking Generalization for Vision Language Action Models”
Source→“ACT — a strong specialist model — drops an average of 28.57% on Single-Arm and 29.06% on Bimanual tasks under visual perturbations like changed lighting, background, object color, or camera pose.”
Source→“ACT, which uses a frozen CLIP language encoder, shows near-zero performance degradation under language paraphrasing: 0.28% drop on Single-Arm, -0.57% on Bimanual.”
Source→“the sim-to-real validation shows an average R² of 0.798 for predicting perturbation-induced success rate changes and a Spearman correlation of 0.916 for rank-ordering perturbation difficulty.”
Source→“Gaurav Sukhatme — USC Department of Computer Science. Senior author and lab director. One of the most prominent figures in robot learning systems research.”
Source→“Jesse Thomason — USC Department of Computer Science. Co-author with expertise in language-conditioned robot policies. Prior work from Thomason's group (Anwar et al. 2024) first documented the catastrophic failure of language generalization under paraphrasing.”
Source→“Ishika Singh — USC Department of Computer Science. Lead researcher on the Colosseum V2 project and co-developer of the original Colosseum benchmark.”
Source→“Unlike existing cloud evaluation systems that focus on robotic arms or simulated agents, ours is the first platform designed for humanoid, which aims to lower the barrier to fair comparison and fosters collaborative progress in humanoid robotics.”
Source→“This paper is a dataset and infrastructure paper, not a breakthrough in robot capabilities. But it matters strategically: it establishes the first large-scale, multimodal, task-diverse benchmark for full-body humanoid manipulation.”
Source→“All end-to-end imitation policies struggle in humanoid manipulation tasks due to the high-dimensional action space in our dataset (which is 28 DoFs in total).”
Source→“The operator wears an Apple Vision Pro to capture wrist and finger keypoints using its cameras on the bottom.”
Source→“Yue Wang is also supported by a Powell Research Award.”
Source→“People get used to this idealized, very predictable form of language, and even people who are not using it, in order to have that sense of powerful, influential writing, they start writing more like LLMs.”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.