robometer

robometer / robometer

Public

Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons

robometer.github.io
31
2
100% credibility
Found Mar 04, 2026 at 31 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Robometer trains AI models to score robot performance in videos, helping distinguish successful, failed, and partial task attempts across diverse robotics scenarios.

How It Works

1
🔍 Discover Robometer

You find Robometer, a friendly tool that watches robot videos and figures out what's going well or going wrong in tasks like picking up objects.

2
📥 Get it ready

Download the program to your computer and follow simple steps to set everything up, like installing a new app.

3
🎥 Test on your videos

Upload videos of your robot trying everyday jobs, like stacking blocks or pouring water, and instantly see scores for progress and success.

4
Choose your style
✅
Use ready helper

Jump right in with the built-in smarts for quick robot feedback.

📚
Teach it yours

Add your robot's videos to make a custom coach tailored just right.

5
🚀 Share your coach

Put your robot coach online so friends or other projects can use it too.

😊 Smarter robots

Your robots now learn from videos what success looks like, trying harder and getting better at real tasks.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 31 to 31 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 robometer?

Robometer is a Python toolkit for scaling general-purpose robotic reward models via trajectory comparisons. It trains models on video trajectories and task descriptions to predict frame-level progress and inter-trajectory preferences, turning expert demos, failures, and suboptimal rolls into usable supervision. Developers get pretrained 4B-parameter models on Hugging Face, plus an HTTP inference server for scoring custom robot videos like "stack the cup."

Why is it gaining traction?

Unlike progress-only models that falter on failures, Robometer blends intra-trajectory progress with preference comparisons for robust, general-purpose rewards across embodiments. Prebuilt evals against baselines (reward alignment, policy ranking) and 1M+ trajectory RBM-1M dataset make benchmarking straightforward. LoRA fine-tuning and dataset converters for 30+ sources lower the barrier for real robotics data.

Who should use this?

Robotics researchers building RLHF-style reward models for manipulation; RL engineers needing video-based progress signals without dense labels; teams with mixed-quality trajectories from sim (LIBERO) or real robots (AgiBotWorld, robomaster setups).

Verdict

Strong foundation from recent arXiv work and HF models/datasets, but 19 stars and 1.0% credibility reflect early maturity—docs are solid, but expect iteration. Prototype your reward pipeline here; production users should validate on held-out tasks.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.