Ashley Edwards

Ashley Edwards is a reinforcement learning researcher whose work includes learning actions from observations and building agents that operate across tasks. She contributed to the concept and project leadership of Genie, as well as its action model research, agent training and demonstration. Her earlier work includes an Atari-capable prototype of Gato and a method for imitating behavior without expert action labels.[2][4][6][7]

The 2024 Genie paper describes a model trained on unlabelled Internet videos that generates environments users can control frame by frame. Edwards's responsibilities spanned both the world model and its use for training agents: the paper credits her with “genie concept, project leadership, action model research, agent training, model demo.”[5][4]

Genie combines a spatiotemporal video tokenizer, an autoregressive dynamics model and a latent action model. Together, these support action control even though the training videos contain no ground-truth action labels. The authors report an 11-billion-parameter model whose learned latent action space also enables agents to imitate behaviors from unseen videos. This connects Edwards's action model research with the project's agent-training work.[5][4]

Edwards discussed Genie on TWIML's August 5, 2024 episode, which identified her affiliation as Runway ML. Latent Space's December 2024 ICML compilation also includes her presentation, “Learning actions, policies, rewards, and environments from videos alone.”[8][1]

Before Genie, Edwards worked on a generalist policy at DeepMind and on imitation learning from observations. These earlier projects provide context for her contributions to agent architectures and latent action learning.[2][6][7][4]

At ICML 2019, Edwards co-authored Imitating Latent Policies from Observation with Himanshu Sahni, Yannick Schroecker and Charles Isbell. Her personal research page lists the same work under the publishing name Ashley D. Edwards.[7][3]

The method infers latent policies directly from state observations by modeling the effects of latent actions and predicting their likelihood. It then uses a small amount of interaction with the environment to map those latent actions to real controls. This alignment permits imitation without expert action labels, while still requiring environment interaction. The authors evaluated it in classic control environments and a platform game; the paper links its code.[7]

Edwards joined Uber AI Labs as a research scientist in March 2019 and DeepMind as a research scientist in September 2020.[2]

For Gato, published in November 2022, Edwards contributed to the first prototype that worked on Atari and explored alternative network architectures and training objectives. The broader team's model used one network with the same weights for tasks including Atari play, image captioning, chat and robotic block stacking. Its context determined whether it produced text, joint torques, button presses or other tokens—giving Edwards's prototype work a place within a policy designed to span modalities and embodiments.[6]