Project Overview

Generative AI models excel at teaching robots complex manipulation skills, but standard models are slow because they generate movements through 10 to 50 repeated calculation steps. This project advances robot learning by adopting Flow Matching instead of traditional diffusion, creating straighter, more dependable motion paths. To make the robot react in real time, MeanFlow is implemented to eliminate step-by-step delays and generate complete robot action trajectories in a single instant step.

The Challenge

Applying generative AI to real-time robotics involves balancing movement quality with speed:

• The Multi-Step Speed Bottleneck: Conventional diffusion policies take 10 to 50 iterative steps to plan a single movement. In robotics, this delay makes the robot sluggish and unable to adapt quickly to changing environments.
• Curved, Complex Paths: Traditional diffusion models follow curved, indirect mathematical paths from random noise to final actions, which can introduce instability during training and execution.

The Solution

To make robot movement both faster and more reliable, I developed an improved generative trajectory system:

• Straighter, More Reliable Movement (Flow Matching): Replaced traditional diffusion with Flow Matching, which guides the robot from random starting noise to desired actions along straight, predictable paths. This makes the robot's behavior noticeably more robust.
• Instant 1-Step Action Generation (MeanFlow): Implemented MeanFlow to bypass repeated iterative loops, enabling the policy to generate full trajectory plans in a single forward pass without losing motion quality.
• Hands-On Benchmark Testing: Validated the models on the Push-T manipulation task, systematically evaluating speed, trajectory smoothness, and task performance under different operating conditions.

Push-T Trajectory Generation Video Demonstration

Benchmark Metrics & Success Rates

Evaluating policy performance across 100 simulation trials per configuration on the Push-T task proved that reducing the generation process from 10 steps down to a single step achieves fast and effective robotic manipulation:

Technologies