I am a Machine Learning Research Engineer at Amazon Web Services and Embodied AI lead at Cohere Labs, and collaborate with PRIOR at Ai2 and the Stanford Geometric Computation Group.
My research interests are in reinforcement learning, computer vision, multimodal foundation models, and scalable approaches for robot learning. My experience ranges from algorithm development to large-scale training and deployment.
j.cole.harrison [at] me [dot] comGoogle ScholarGitHubXLinkedIn
Papers
News
- [] Co-organized the first CVPR workshop on Embodied Reasoning in Action (ERA) at CVPR 2026.
- [] Challenge organizer for the 1st ICCV Workshop on Category-Level Object Pose Estimation in the Wild (WCLOP).
Code
-
TOPReward
Reference implementation and evaluation tooling for extracting dense task-progress signals from pretrained video-language models.
-
MolmoAct 2
Pretraining data-quality filtering using TOPReward thresholding, and value-head work, for an open action-reasoning model family released with models, data, and evaluation rollouts.
-
UniPose9D
Implementation and pretrained models for category-agnostic 9D object pose estimation from RGB and RGB-D input, with no mesh, category label, or reference view required.
-
LeRobot
TOPReward and MolmoAct 2 integrations into Hugging Face's robotics library, plus fixes for Koch 1.1 hardware and setup failure modes.
-
MeGPT
A QLoRA fine-tuning pipeline for training a personalized Llama-based chatbot from private message history.