RESEARCH PAPER

DynaMo: In-Domain Dynamics Pretraining for Visuo-Motor Control

Zichen Jeff Cui; Hengkai Pan; Aadhithya Iyer; Siddhant Haldar; Lerrel Pinto

Classification

View four quadrants
Major category
Related resources
Architecture
Not applicable
Prediction paradigm
Not applicable
Source review status
Not assigned

Category review. Action-free inverse/forward dynamics pretrain visual features used by separate visuomotor policies. This is relevant representation-method context for WAMs, rather than a canonical pretrained backbone or a deployed future-prediction controller. The paper is retained for its adjacent resource/method role; WAM joint-prediction/IDM architecture quadrants do not apply to the cataloged contribution. Reading evidence

AT A GLANCE

Contribution

DynaMo uses action-free dynamics pretraining to make visual features useful for imitation learning. An encoder, latent inverse model and forward model jointly learn to predict next-frame embeddings; a separate policy then learns from frozen features and labeled demonstrations. Results favor DynaMo on several manipulation tasks, but include ties and initialization regressions. The contribution is a representation-learning objective, with no demonstrated use of the dynamics models for online planning.

Abstract

An abstract has not been added yet.

Affiliations

New York University