yzc0731

yzc0731 / HinFlow

Public

Official Code Implementation of Translating Flow to Policy via Hindsight Online Imitation

115
0
100% credibility
Found Feb 17, 2026 at 87 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

HinFlow is a research codebase for training robots to perform manipulation tasks by learning from video demonstrations using hindsight imitation techniques on benchmarks like LIBERO and ManiSkill.

How It Works

1
🔍 Discover smart robot trainer

You find this tool that teaches robots household tasks by watching example videos, like picking up bowls or opening drawers.

2
📥 Gather learning videos

Download ready-made video clips of robots doing everyday chores from a shared collection.

3
🛠️ Prepare the videos

Smooth out the videos and mark key points to help the robot learn movements.

4
🧠 Train the big-picture thinker

Teach the robot to plan overall steps, like 'go to the drawer first', using the prepared videos.

5
💪 Train precise hand moves

Fine-tune the robot's actions to match real movements, building on the plans.

6
🎯 Test on new chores

Run the robot on fresh tasks and compare with simple methods to see improvements.

✅ Robots master tasks

Your robot now handles picking, placing, and more just from watching videos, ready for real homes.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 87 to 115 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 HinFlow?

HinFlow translates optical flow point tracks from video demos into executable robot policies using hindsight online imitation learning. It handles preprocessing with CoTracker, trains a high-level planner on subgoals, then a low-level policy on pixel observations and tracks from dual cameras (agentview and eye-in-hand). Python-based with PyTorch Lightning, it supports LIBERO and ManiSkill benchmarks via Hugging Face datasets and conda setup for quick repro of ICLR 2026 paper results.

Why is it gaining traction?

This official GitHub repo stands out for beating baselines like BC and ATM on long-horizon manipulation without proprioception, using just RGB+tracks. Developers dig the two-stage pipeline—planner checkpoints on HF, policy training with action chunking and augmentation—for reliable sim-to-real potential. Scripts for data collection, labeling, and eval make experimenting on tasks like microwave opening or cube poking straightforward.

Who should use this?

Robotics PhD students or researchers benchmarking vision-based imitation on LIBERO suites (book, butter, chocolate, microwave) or ManiSkill (place-sphere, poke-cube, pull-cube-tool). Ideal for teams prototyping flow-to-policy without custom trackers, especially if you're tired of proprio-heavy methods failing on occlusions.

Verdict

Grab it if you're in robot learning—solid repro scripts and HF integration make it low-risk to test, despite 87 stars and 1.0% credibility signaling early maturity. Docs are paper-focused; expect some env tweaks for your setup.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.