Train a robot arm to perform long multi-step tasks from human demonstration videos.
Benchmark your own imitation learning method against Chronos on the RMBench simulation.
Collect demonstration data from real dual-arm UR3 robots and train the Chronos model on it.
Run a trained Chronos model on a physical robot so it reacts to camera input in real time.
| yulinzhouzyl/chronos | andreanocalvin/autoclaw-autologin | cortex-trading-systems/polymarket-copy-trading-bot-clob-ai | |
|---|---|---|---|
| Stars | 51 | 51 | 51 |
| Language | Python | Python | Python |
| Setup difficulty | hard | moderate | hard |
| Complexity | 5/5 | 4/5 | 3/5 |
| Audience | researcher | developer | general |
Figures from each repo's GitHub metadata at analysis time.
Requires real or simulated robot hardware, camera setup, and deep-learning training infrastructure, still in active development.
{"i": 0, "repo": "yulinzhouZYL/Chronos", "stars": 51, "lang": "Python", "ok": true, "char_count": 1983, "text": "Chronos is a research project from a team working on robotics and imitation learning. The goal is to teach robot arms to perform complex, long sequences of actions by learning from human demonstrations. A key challenge in this area is that robots often forget what they were doing earlier in a task. Chronos addresses this by keeping a full memory of past observations and using a physics-inspired mathematical method to plan smooth, realistic movements.\n\nThe repository provides code for two types of experiments. The first is a simulation environment called RMBench, which lets researchers test the system on virtual tasks like covering blocks. The second involves real physical robots, specifically dual-arm UR3 setups. The real-world code includes everything needed to collect demonstration data, train the model, and run the robot in a closed loop where it reacts to what it sees.\n\nIn their experiments, the authors report strong results. On the RMBench benchmark, Chronos achieved a 73.6% average success rate, outperforming a compared method called Mem-0 by 22.8 percentage points while using 30 times fewer parameters. In four real-world dual-arm tasks, Chronos reached a 78% success rate, compared to 7% for another method called pi0.5.\n\nThe codebase is organized into separate folders for simulation and real-world work. The simulation folder contains scripts for collecting data, computing normalization statistics, training the model, and running evaluations. The real-world folder provides code for handling camera inputs, robot control, data collection, and inference. The project also shares pre-trained model checkpoints on Hugging Face, a platform for hosting model files.\n\nThe README is detailed, covering installation steps and setup instructions. The authors note that additional benchmark code for ALOHA and RoboTwin 2.0 will be released later, so the repository is still in active development.", "error": null}
Chronos is a research project that teaches robot arms to perform long, complex tasks by learning from human demonstrations and using a physics-inspired memory method so robots remember what they were doing.
Mainly Python. The stack also includes Python, PyTorch, UR3 robot arms.
No license information is provided, so rights to use or modify this code are unclear.
Setup difficulty is rated hard, with roughly 1day+ to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
double-check against the repo, no cap.