rise-robot

name	rise-robot
title	RISE: Self-Improving Robot Policy with Compositional World Model
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2602.11075
keywords	["Robot Learning","World Models","Reinforcement Learning","Imagination","Policy Optimization","Compositional"]
description	Enable robot policies to self-improve through imagination using learned dynamics and value models without physical trial-and-error. Compositional world model separates concerns enabling 35-45% performance gains on contact-rich manipulation.

RISE: Self-Improving Robot Policy with Compositional World Model

Problem Context

Vision-Language-Action (VLA) foundation models excel at semantic understanding but struggle with precise contact-rich manipulation requiring fine motor control. Physical trial-and-error RL is expensive and impractical for real robots. Standard world models struggle to balance action controllability with realistic prediction. The core challenge: learn manipulation policies efficiently without costly physical interaction.

Core Concept

RISE shifts robot learning from the physical world to imagination using a compositional world model with specialized dynamics and value components. Rather than joint training, the approach separates:

1. Dynamics Model: Predicts future observations given actions (what happens)
2. Value Model: Evaluates imagined states for progress signals (is this good)

This separation enables on-policy RL in imagination, where the policy iteratively improves by proposing actions, simulating consequences, evaluating outcomes, and updating based on advantage signals—all without touching a physical robot.

Architecture Overview

• Dynamics Model: Video diffusion architecture with Task-Centric Batching for action diversity
• Value Model: Dual-objective learning combining progress regression and Temporal-Difference signals
• Advantage Computation: Cumulative improvement across imagined trajectories
• Policy Warm-up: Offline initialization on real demonstrations
• Self-Improving Loop: Imagination-based RL for continuous refinement
• Multi-view Setup: Observations from multiple camera perspectives for robustness

Implementation

The compositional world model operates through two specialized components trained with different objectives:

class DynamicsModel(nn.Module):
    """Predicts future observations from current state and action."""

    def __init__(self):
        # Video diffusion backbone
        self.backbone = VideoDiffusion()

    def forward(self, observations, actions):
        
        
        state_action = torch.cat([observations, actions], dim=)
        
        predictions = .backbone.sample(state_action, steps=)
         predictions