Lirui Wang, Yiyang Ling, Zhecheng Yuan, Mohit Shridhar, Chen Bao, Yuzhe Qin, Bailin Wang, Huazhe Xu, Xiaolong Wang
Project Page | Arxiv | Gradio Demo | Huggingface Dataset | Finetuned Code-LLama Model | GPTs
This repo explores the use of an LLM code generation pipeline to write simulation environments and expert goals to augment diverse simulation tasks. Strongly recommend also checking out the Gradio Demo and GPTs.
pip install -r requirements.txt
python setup.py develop
export GENSIM_ROOT=$(pwd)
export OPENAI_KEY=YOUR KEY
. We use OpenAI's GPT-4 as the language model. You need to have an OpenAI API key to run task generation with GenSim. You can get one from here.
After the installation process, you can run:
# basic bottom-up prompt
python gensim/run_simulation.py disp=True prompt_folder=vanilla_task_generation_prompt_simple
# bottom-up template generation
python gensim/run_simulation.py disp=True prompt_folder=bottomup_task_generation_prompt save_memory=True load_memory=True task_description_candidate_num=10 use_template=True
# top-down task generation
python gensim/run_simulation.py disp=True prompt_folder=topdown_task_generation_prompt save_memory=True load_memory=True task_description_candidate_num=10 use_template=True target_task_name="build-house"
# task-conditioned chain-of-thought generation
python gensim/run_simulation.py disp=True prompt_folder=topdown_chain_of_thought_prompt save_memory=True load_memory=True task_description_candidate_num=10 use_template=True target_task_name="build-car"
- To remove a task (delete its code and remove it from the task and task code buffer), use
python misc/purge_task.py -f color-sequenced-block-insertion
- To add a task (extract task description to add to buffer), use
python misc/add_task_from_code.py -f ball_on_box_on_container
- All generated tasks in
cliport/generated_tasks
should have automatically been imported - Set the task name and then use
demo.py
for visualization. For instance,python cliport/demos.py n=200 task=build-car mode=test disp=True
. - The following is a guide for training everything from scratch (More details in cliport). All tasks follow a 4-phase workflow:
- Generate
train
,val
,test
datasets withdemos.py
- Train agents with
train.py
- Run validation with
eval.py
to find the best checkpoint onval
tasks and save*val-results.json
- Evaluate the best checkpoint in
*val-results.json
ontest
tasks witheval.py
- Generate
-
Prepare data using
python gensim/prepare_finetune_gpt.py
. Released dataset is here -
Finetune using openai api
openai api fine_tunes.create --training_file output/finetune_data_prepared.jsonl --model davinci --suffix 'GenSim'
-
Evaluate it using
python gensim/evaluate_finetune_model.py +target_task=build-car +target_model=davinci:ft-mit-cal:gensim-2023-08-06-16-00-56
-
Compare with
python gensim/run_simulation.py disp=True prompt_folder=topdown_task_generation_prompt_simple load_memory=True task_description_candidate_num=10 use_template=True target_task_name="build-house" gpt_model=gpt-3.5-turbo-16k trials=3
-
Compare with
python gensim/run_simulation.py disp=True prompt_folder=topdown_task_generation_prompt_simple_singleprompt load_memory=True task_description_candidate_num=10 target_task_name="build-house" gpt_model=gpt-3.5-turbo-16k
-
turbo finetuned models.
python gensim/evaluate_finetune_model.py +target_task=build-car +target_model=ft:gpt-3.5-turbo-0613: trials=3 disp=True
-
Finetune Code-LLAMA using hugging-face transformer library here
-
offline eval:
python -m gensim.evaluate_finetune_model_offline model_output_dir=after_finetune_CodeLlama-13b-Instruct-hf_fewshot_False_epoch_10_0
- Note that the 100+ generated tasks by GenSim can be used for benchmarking algorithms in multitask policy training. See
scripts/task_list/GPT_*.json
for a list of benchmark settings. Pretrained multitask models can be found here. - Generate multitask demonstrations. For example, run
bash scripts/generate_datasets.sh data 'align-box-corner assembling-kits block-insertion'
- Single-task training
sh scripts/train_test_multi_task.sh data "[align-rope,align-box-corner]
- Multi-task training
sh scripts/train_test_single_task.sh data align-box-corner
- Temperature
0.5-0.8
is good range for diversity,0.0-0.2
is for stable results. - The generation pipeline will print out statistics regarding compilation, runtime, task design, and diversity scores. Note that these metric depend on the task compexity that LLM tries to generate.
- Core prompting and code generation scripts are in
gensim
and training and task scripts are incliport
. prompts/
folder stores different kinds of prompts to get the desired environments. Each folder contains a sequence of prompts as well as a meta_data file.prompts/data
stores the base task library and the generated task library.- The GPT-generated tasks are stored in
generated_tasks/
. Usedemo.py
to play with them.cliport/demos_gpt4.py
is an all-in-one prompt script that can be converted into ipython notebook. - Raw text outputs are saved in
output/output_stats
, figure results saved inoutput/output_figures
, policy evaluation results are saved inoutput/cliport_output
. - To debug generated code, manually copy-paste
generated_task.py
then runpython cliport/demos.py n=50 task=gen-task disp=True
- This version of cliport should support
batchsize>1
and can run with more recent versions of pytorch and pytorch lightning. - Please use Github issue tracker to report bugs. For other questions please contact Lirui Wang
- blender rendering
python cliport/demos.py n=310 task=align-box-corner mode=test disp=True +record.blender_render=True record.save_video=True
If you find GenSim useful in your research, please consider citing:
@inproceedings{wang2023gen,
author = {Lirui Wang and Yiyang Ling and Zhecheng Yuan and Mohit Shridhar and Chen Bao and Yuzhe Qin and Bailin Wang and Huazhe Xu and Xiaolong Wang},
title = {GenSim: Generating Robotic Simulation Tasks via Large Language Models},
booktitle = {Arxiv},
year = {2023}
}