Whole-Body Conditioned Egocentric Video Prediction

Predicting Ego-centric Video from human Actions (PEVA). Given past video frames and an action specifying a desired change in 3D pose, PEVA predicts the next video frame. Our results show that, given the first frame and a sequence of actions, our model can generate videos of atomic actions (a), simulate counterfactuals (b), and support long video generation (c).

Recent years have brought significant advances in world models that learn to simulate future outcomes for planning and control. From intuitive physics to multi-step video prediction, these models have grown increasingly powerful and expressive. But few are designed for truly embodied agents. In order to create a World Model for Embodied Agents, we need a real embodied agent that acts in the real world. A real embodied agent has a physically grounded complex action space as opposed to abstract control signals. They also must act in diverse real-life scenarios and feature an egocentric view as opposed to aesthetic scenes and stationary cameras.

Why It's Hard

Action and vision are heavily context-dependent. The same view can lead to different movements and vice versa. This is because humans act in complex, embodied, goal-directed environments. Human control is high-dimensional and structured. Full-body motion spans 48+ degrees of freedom with hierarchical, time-dependent dynamics. Egocentric view reveals intention but hides the body. First-person vision reflects goals, but not motion execution, models must infer consequences from invisible physical actions. Perception lags behind action. Visual feedback often comes seconds later, requiring long-horizon prediction and temporal reasoning.

To develop a World Model for Embodied Agents, we must ground our approach in agents that meet these criteria. Humans routinely look first and act second—our eyes lock onto a goal, the brain runs a brief visual “simulation” of the outcome, and only then does the body move. At every moment, our egocentric view both serves as input from the environment and reflects the intention/goal behind the next movement. When we consider our body movements, we should consider both actions of the feet (locomotion and navigation) and the actions of the hand (manipulation), or more generally, whole-body control.

What Did We Do?

We trained a model to Predict Ego-centric Video from human Actions (PEVA) for Whole-Body-Conditioned Egocentric Video Prediction. PEVA conditions on kinematic pose trajectories structured by the body’s joint hierarchy, learning to simulate how physical human actions shape the environment from a first-person view. We train an autoregressive conditional diffusion transformer on Nymeria, a large-scale dataset pairing real-world egocentric video with body pose capture. Our hierarchical evaluation protocol tests increasingly challenging tasks, providing comprehensive analysis of the model’s embodied prediction and control abilities. This work represents an initial attempt to model complex real-world environments and embodied agent behaviors through human-perspective video prediction.

Method

Structured Action Representation from Motion

To bridge human motion and egocentric vision, we represent each action as a rich, high-dimensional vector capturing both full-body dynamics and detailed joint movements. Instead of using simplified controls, we encode global translation and relative joint rotations based on the body’s kinematic tree. Motion is represented in 3D space with 3 degrees of freedom for root translation and 15 upper-body joints. Using Euler angles for relative joint rotations yields a 48-dimensional action space (3 + 15 × 3 = 48). Motion capture data is aligned with video using timestamps, then converted from global coordinates to a pelvis-centered local frame for position and orientation invariance. All positions and rotations are normalized to ensure stable learning. Each action captures inter-frame motion changes, enabling the model to connect physical movement with visual consequences over time.

Design of PEVA: Autoregressive Conditional Diffusion Transformer

While the Conditional Diffusion Transformer (CDiT) from Navigation World Models uses simple control signals like velocity and rotation, modeling whole-body human motion presents greater challenges. Human actions are high-dimensional, temporally extended, and physically constrained. To address these challenges, we extend the CDiT method in three ways: random timeskips, sequence-level training, and action embeddings.

Sampling and Rollout Strategy

At test time, we generate future frames by conditioning on a set of past context frames. We encode these frames into latent states and add noise to the target frame, which is then progressively denoised using our diffusion model. For action-conditioned prediction, we use an autoregressive rollout strategy. Starting with context frames, we encode them using a VAE encoder and append the current action. The model then predicts the next frame, which is added to the context while dropping the oldest frame, and the process repeats for each action in the sequence. Finally, we decode the predicted latents into pixel-space using a VAE decoder.

Planning and Visual Ability

PEVA can be used for planning by simulating multiple action candidates and scoring them based on their perceptual similarity to the goal, as measured by LPIPS. We formulate planning as an energy minimization problem and perform action optimization using the Cross-Entropy Method (CEM). This allows the model to predict sequences of actions that lead to specific goals, such as reaching for a kettle or finding a path to the fridge.