linhanwang

linhanwang / Drive-JEPA

Public

Repo of "Drive-JEPA: Video JEPA Meets Multimodal Trajectory Distillation for End-to-End Driving"

71
2
69% credibility
Found Feb 05, 2026 at 45 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

Drive-JEPA is a research project that combines learning from driving videos with smart path planning to boost performance in autonomous driving systems.

How It Works

1
🔍 Discover Drive-JEPA

You come across this project while exploring new ideas for smarter self-driving cars.

2
📖 Read the big idea

You learn how it uses driving videos to help cars predict and plan safer paths ahead.

3
🌟 See the master plan

You get excited viewing the clear steps it takes to learn from videos and choose the best driving routes.

4
👀 Check the visuals

You watch examples showing how it creates several smart options for driving, making choices more flexible and safe.

5
📋 Look at what's next

You note the plans to share ready-to-use tools and examples soon.

🎉 Feel the future

You're inspired by this step toward safer, more reliable self-driving technology.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 45 to 71 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 Drive-JEPA?

Drive-JEPA brings Video JEPA self-supervised pretraining to end-to-end autonomous driving, distilling multimodal trajectories from simulators alongside human data for robust planning. In Python, it processes driving videos into predictive features via a ViT encoder, then generates diverse trajectory proposals scored for safety and progress. Users download datasets/maps with bash scripts, cache features, train on L40S GPUs, and eval on NAVSIM benchmark—pretrained weights on Hugging Face speed things up.

Why is it gaining traction?

It crushes NAVSIM scores (93.3 PDMS v1 perception-based, 87.8 EPDMS v2) by avoiding unimodal human traj pitfalls through distillation, enabling stable multimodal behaviors. Perception-free mode still beats priors by 3 PDMS; setup mirrors NAVSIM with env vars and one-liner HF downloads. Devs dig the momentum-aware selection for comfy, safe drives over brittle alternatives.

Who should use this?

AV researchers benchmarking planners on NAVSIM/nuPlan sims, especially those blending video world models with trajectory optimization. Teams prototyping perception-based agents without lidar reliance. Sim hackers tweaking distillation for custom driving data.

Verdict

Solid SOTA starter for AV sim work—train/eval flows work out-of-box—but 50 stars and 0.7% credibility score mean docs are README-thin, tests sparse; fork and debug expected. Use if NAVSIM's your jam.

(187 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.