Guang000

Guang000 / RC-aux

Public

Official Code of "Predictive but Not Plannable: RC-aux for Latent World Models"

guang000.github.ioRC-aux-Webpage
38
1
89% credibility
Found May 20, 2026 at 38 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

RC-aux is a research project that helps robots learn to predict and plan better. It builds on world models that learn to imagine what happens next when a robot acts, and adds a special component that learns whether a goal can be reached within a certain number of steps. This extra knowledge helps robots plan more accurately and succeed at tasks like pushing objects, reaching targets, and navigating through rooms. The project includes training code, evaluation tools, and support for various robotic manipulation tasks using video-based learning.

How It Works

1
📚 Discover the research project

You come across the paper while reading about robot learning and world models, and it catches your attention because it addresses a key problem: making predictions about what robots can actually achieve.

2
🖥️ Set up the environment

You download the code and install it on your computer. The setup includes everything needed for training robot control models from images.

3
🎓 Prepare your training data

You organize video demonstrations of robot tasks—like pushing objects, reaching goals, or navigating through rooms—into a format the system can use.

4
🧠 Train the model to predict futures

You let the system learn from the demonstrations. The model learns to imagine what will happen when it takes different actions, and it also learns which goals are actually reachable in a given number of steps.

5
Apply the model to different robots
🎯
Real-time planning

The model thinks ahead and chooses actions by imagining multiple possible futures, picking the path most likely to reach the goal.

⚡
Action prediction

A faster system that has learned from your model to directly predict good actions without explicit planning.

🎉 Watch your robot complete tasks

Your robot successfully reaches goals across multiple challenges—from simple reaching tasks to complex manipulation—thanks to the improved prediction and planning capabilities.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 38 to 38 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 RC-aux?

RC-aux is a lightweight training objective that fixes a key weakness in latent world models: they predict future states well but struggle to plan actions toward goals. The system adds a reachability head that learns whether a robot can reach a goal within a given time budget, then uses that signal to weight planning costs. It drops into existing LeWM-based systems without changing the core model architecture. The Python codebase trains with PyTorch Lightning and supports MPC planning with CEM and gradient-based solvers across five pixel-control benchmarks: TwoRoom, Reacher, Push-T, Cube, and Wall environments.

Why is it gaining traction?

The paper shows meaningful gains on challenging tasks, especially the Wall environment where RC-aux jumps from 50% to 83% success rate. The approach is modular: you can continue training from an existing LeWM checkpoint rather than training from scratch. Configurable reachability cost weights let practitioners tune the planning sensitivity without retraining. The system handles both robot manipulation (via the LIBERO-Goal benchmark extension) and classic control tasks, making it a single codebase for multiple domains.

Who should use this?

Researchers working on model-based reinforcement learning who want to improve goal-conditioned planning without overhauling their world model. Robotics teams evaluating latent representations for manipulation tasks will find the LIBERO-Goal scripts directly applicable. This is not production-ready for real robot deployment; it targets offline evaluation and algorithm research with MuJoCo-based simulation.

Verdict

With a 0.9% credibility score and only 38 stars, this is an early-stage research project from academia -- solid paper, limited community validation. The documentation is thorough for reproducing paper results, but test coverage is minimal and there's no pip package for easy installation. Clone the repo if you're extending world model planning research; wait for wider adoption before building products on top of it.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.