Many thanks to my co-authors Gen Li and @SiyuTang3. The paper, along with additional video results and code, are available at: . 8/8.

In addition to the noise optimization-based control, we propose a reinforcement learning-based method for fast online generation. It models motion control as a Markov decision process with a latent action space and use RL to learn efficient goal-achieving policies. 7/8

What do you want to know?.

The first method employs gradient descent optimization over latent diffusion noise variables to find optimal solutions. This plug-and-play approach supports diverse spatial control tasks, like keyframe body or joint trajectory-based control, as shown in the video results. 6/8

To eliminate implausible motion sequence solutions and enhance generation realism, we propose two methods that leverage the latent space learned by the motion primitive diffusion model to find high-quality motion sequence solutions. 5/8.

While language provides a human-friendly interface for controlling semantics, relying solely on text fails us in precise spatial control. We formulate motion control as a minimization problem:.4/8

DartControl uses an architechture of autoregressive latent diffusion to learn a motion primitive space jointly conditioned on the text prompts and motion history. The trained denoiser and decoder models enable real-time, text-driven motion generation. 3/8

Each frame of the human motion primitives is represented using an overparamterized representation that supports real-time animation and spatial control tasks. 2/8

DartControl decomposes long motion sequences into overlapping motion primitives, which more tracable for generative learning and suitable for online generation. 1/8

#ICLR2025.Introducing DartControl, a diffusion-based autoregressive motion model for real-time text-driven motion control. Moreover, it enbales various motion generation tasks with spatial goals. Join us at Hall 3+2B #94 on April 24, 3–5:30 p.m!.Website:

Method: .DIMOS formulates synthesizing human motions in 3D scenes as a Markov decision process with latent action spaces, which are learned from motion capture data. The virtual human agents are driven by scene-aware and goal-driven policies to synthesize various behaviors. 3/3

Given a 3D scene, DIMOS enables virtual humans to navigate in the environment and interact with objects in a realistic manner. Furthermore, DIMOS can generate more diverse human motions than previous methods, reflecting the subtle differences in real human behaviors. 2/3.

#ICCV2023 Introducing DIMOS: Synthesizing Diverse Human Motions in 3D Indoor Scenes. All info about DIMOS: Catch our poster presentation in Paris at #ICCV2023:.📅 Date: Thursday, 5th .🕒 Time: 14:30 - 16:30 .📍 Location: Room "Foyer Sud" 072. 1/3

RT @jbhuang0604: How to organize your talk?. Qual/prelim/defense/conference season is coming up! 😱 How should we organize the talk so that….