GigaAI-research

GigaAI-research / ViVa

Public

ViVa: A Video-Generative Value Model for Robot Reinforcement Learning

45
1
100% credibility
Found Apr 15, 2026 at 46 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

ViVa trains a video model to predict robot task success from camera views and robot states, aiding reinforcement learning.

How It Works

1
🔍 Discover ViVa

You find this robot helper project while reading about smarter robot learning.

2
🛠️ Set up your workspace

Create a simple space on your computer to prepare everything needed.

3
📥 Gather robot videos and tools

Download example robot task videos and ready-made brain parts for your helper.

4
🤖 Train your robot coach

Feed the videos to teach it how robots succeed at tasks like stacking or folding.

5
🎥 Test on new robot actions

Show it fresh robot videos and ask it to guess task progress.

✅ See clear progress insights

Watch visualizations revealing exactly how well the robot is doing toward success.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 46 to 45 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 ViVa?

ViVa is a Python toolkit that adapts pretrained video diffusion models into value estimators for robot reinforcement learning. Feed it multi-view camera images and robot proprioception from datasets like LeRobot, and it outputs predicted future joint states plus a scalar score tracking task progress to success. It tackles partial observability and sparse rewards in real-world manipulation, boosting setups like box assembly when plugged into actor-critic frameworks.

Why is it gaining traction?

This open viva github project skips training video models from scratch by recycling spatiotemporal priors from giants like WAN text-to-video generators, enabling reliable long-horizon value signals that generalize to unseen objects. Users dig the streamlined workflow: grab WAN weights via Hugging Face, precompute T5 task embeddings, train multi-task on 8 GPUs, then batch-infer predictions directly into LeRobot parquet files for policy fine-tuning. No fuss with custom diffusion heads—pure plug-and-play for robot learning models.

Who should use this?

Robot RL engineers wrangling LeRobot datasets for manipulation tasks like cloth folding or toilet paper handling. Teams building VLA policies that need grounded value functions amid camera noise and delayed feedback. Hardware roboticists evaluating progress on real arms without manual labeling.

Verdict

Promising for robot RL experimentation—solid arXiv paper, working quickstart, multi-GPU inference shines on beefy setups—but 45 stars and 1.0% credibility score scream early days; docs are README-focused with no tests. Fork the github viva project and validate on your data before betting the farm.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.