Update README.md

README.md

@@ -15,28 +15,34 @@ license: apache-2.0
 
 This demo is from the paper:
 
-<!-- [Code as Policies: Language Model Programs for Embodied Control](https://code-as-policies.github.io/)
- -->
-Below is an interactive demo for the simulated tabletop manipulation domain, seen in the paper section IV.D
+GenSim: Supersizing Robotic Simulation Tasks for Policy Learning via Generative Models
+
 
 ## Preparations
 1. Obtain an [OpenAI API Key](https://openai.com/blog/openai-api/)
 
 ## Usage 
-1. Type in desired task name in the box. Then GenSim will try to run through the pipeline
-2. The task name has the form word separated by dash. For instance, 'place-blue-in-yellow' and 'align-rainbow-along-line'.
+0. Click Run-Example will simulate one example of pre-saved tasks in the task library and render videos.
+1. Top-Down Model:
+  0. Type in the desired task name in the box. Then GenSim will try to run through the pipeline to generate the task.
+  1. The task name has the form word separated by dash. For instance, 'place-blue-in-yellow' and 'align-rainbow-along-line'.
+2. Bottom-Up Model: No need to type in desired task. GenSim will try to generate a novel tasks that is different from the task library.
+3. Usage: Always click on "Setup/Reset Simulation" and then click "Run".
 
 ## Guideline
+0. The first output is the current stage of the task generation pipeline.
+1. The second output shows the generated code from Gen-Sim
+2. If there are errors in the generation stage above, you will see an error log on the top right.
+3. If there is still orange bars around the third output, then it means the task is being simulated and rendered.
+4. The rendered video will come out in a stream, i.e. it will render and re-render in a sequence. Each new update takes 15 seconds.
+
 
 ## Known Limitations
-1. The code generation can fail or generate infeasible tasks. 
+1. The code generation can fail or generate infeasible tasks. The success rate is around 0.5. 
 2. The low-level pick place primitive does not do collision checking and cannot pick up certain objects.
 3. Top-down generation is typically more challenging if the task name is too vague or too distant from motions such as stacking.
 
 
 ## Note
-For GPT-4 model, each inference costs about $\\$$0.3. For GPT-3.5 model, each inference costs about $\\$$0.03.
-
+For GPT-4 model, each inference costs about $\\$$0.3. For GPT-3.5 model, each inference costs about $\\$$0.03. You can select which LLM model you would like to use.
 
-## Acknowledgement
-Thanks to Jacky's [code-as-policies](https://huggingface.co/spaces/jackyliang42/code-as-policies/tree/main) demo.