real-stanford

real-stanford / gated-memory-policy

Public

Official Implementation of Paper [Gated Memory Policy], arXiv:2604.18933

gated-memory-policy.github.io
19
0
100% credibility
Found May 03, 2026 at 19 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

A Stanford research repository for training robot AI policies that use gated memory to excel at long-horizon manipulation tasks in simulation and real-world settings.

How It Works

1
🔍 Discover Stanford's robot memory project

You stumble upon this exciting research from Stanford that teaches robots to remember past actions for smarter moves.

2
💻 Set up your training space

Prepare your computer with the right tools so everything runs smoothly.

3
📥 Download robot lessons and skills

Grab pre-made practice videos and ready robot brains from a shared collection.

4
🧠 Train a remembering robot brain

Watch as your robot learns to use memory for tough tasks like picking, pushing, or multi-step games.

5
🎮 Test in a virtual playground

Run trials in simulated worlds to see your robot succeed on benchmarks like cloth flinging or shape matching.

6
Choose your test world
🖥️
Stay in simulation

Perfect skills without hardware risks.

⚙️
Deploy to real robot

Connect to arms like UR5 for hands-on results.

✅ Robot masters memory tasks

Celebrate as your robot handles long challenges, remembering colors, shapes, and sequences flawlessly!

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 19 to 19 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is gated-memory-policy?

This Python repo delivers the official implementation of Gated Memory Policy (arXiv:2604.18933), a memory-gated diffusion policy for robotic imitation learning on long-horizon tasks. It lets you train vision-based policies that selectively remember past observations to handle occlusion and multi-step manipulation, with pretrained models and datasets on Hugging Face. Users get tools for simulation benchmarks like Memimic, RoboMimic, and Mikasa-Robo, plus real-world deployment on UR5 or ARX5 arms via iPhUMI hardware.

Why is it gaining traction?

Unlike basic diffusion policies, its gated memory boosts success on memory-intensive tasks like object tracking under occlusion, with modular components for easy policy serving, batch rollouts, and Hydra-config training. Official GitHub repository structure separates policy training, sim envs, and real-robot setups, plus scripts for downloading 48GB checkpoints or 325GB datasets. Developers dig the zero-shot sim-to-real transfer and multi-GPU eval sweeps.

Who should use this?

Robotics PhD students or engineers benchmarking imitation policies on ManiSkill or RoboSuite tasks like pick-place or tool hanging. Real-robot teams with UR5e grippers needing in-the-wild deployment for cloth flinging or iterative casting. Anyone extending diffusion transformers with memory for 3D-spatial multimodal memory in sparse-reward settings.

Verdict

Worth forking for robotics imitation work—strong docs, HF integration, and real-robot support outweigh the 19 stars and 1.0% credibility score. Still early; test on sim first before hardware commits. (198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.