Loading activity…
@du_yilun
Assistant Professor at Harvard @KempnerInst + CS. PhD @MIT_CSAIL, BS MIT. Generative Models, Compositionality, Embodied Agents, Robot Learning.
AI
The VisualForce team reports 73.3% success on berry picking versus 0% for its baseline, across 15 trials each.
XAI
A post introducing TEM claims better generalization, test-time scaling and out-of-distribution detection than diffusion models.
XAI
A researcher says GTP proposes promising action sequences and refines them with a learned world model, especially for manipulation tasks.
XAI
The researchers report over 2,000× lower error than local regression when recovering dynamics in a double-well system.
XAI
The ScaleMAS announcement says agents publish findings in a shared database and link to work they build on.
AI
A post describing RAX reports a 78.3% RoboDojo result, versus 24.9% for the prior state of the art.
AI
The researchers say the method pairs reconstruction pretraining with distributional post-training, without relying on a diffusion teacher.
AI
A user highlights NeurIPS work on estimating SE(3) transforms from video and suggests a possible robotics use.
AI
The SIMPACT project describes an agent that proposes rough actions, tests them in simulation and refines them before a robot executes them in the real world.
AI
A post links to the paper “Training Object Permanence in World Models” on Hugging Face.
XAI
Harvard assistant professor outlines his introduction of Comet as compositional energy concepts.
AI
Harvard assistant professor states COMET enables wider research adoption.
AI
Researchers on X compare the new Explorative Modeling paper to earlier IMLE methods in algorithms, loss, and motivation.
AI
Side-by-side excerpts highlight similar mode forcing sections in XM and CHIMLE papers.
