search

simpler-env / SimplerEnv

Evaluating and reproducing real-world robot manipulation policies (e.g., RT-1, RT-1-X, Octo) in simulation under common setups (e.g., Google Robot, WidowX+Bridge)
https://simpler-env.github.io/
MIT License
250 stars 28 forks source link

Any way to get expert demos for each task? #31

Open MasterXiong opened 3 weeks ago

MasterXiong commented 3 weeks ago

Hi,

Thanks for open-sourcing this brilliant benchmark! I was wondering that is there any quick way to get expert demos for each task in Simpler? Like a script oracle policy, or a RL policy trained on each env?

I didn't find anything like these in the codebase. But I was wondering if the authors may have run some relevant experiments before and could help share some insights on this? Like how good can a RL policy perform in each env? Thanks!

xuanlinli17 commented 3 weeks ago

The "expert demos" are from the open-x-embodiment dataset.

MasterXiong commented 3 weeks ago

Sorry for the confusion. I mean expert demos for each Simpler env in simulation. I guess directly running a open-x-embodiment action sequence in simpler would just fail due to dynamics mismatch?

xuanlinli17 commented 3 weeks ago

Open-loop unrolling demo trajectories from open-x-embodiment will most likely be successful, since simpler-env already performs sysid to match the dynamics gap between sim & real.

MasterXiong commented 1 week ago

Thanks for your reply! If I want to run a OXE expert trajectory in a open-loop way in the corresponding simpler environment, how should I setup and reset the environment to match with those in the expert trajectory? Thanks!

xuanlinli17 commented 1 week ago

For existing environments in Simpler, you just need to take the trajectories with the same language instruction from the demo dataset and adjust the initial object pose to match the demo's first frame. You can then open-loop unroll the demo trajectory.

For environments not already present in this repo, you need to create new environment, which means you need the assets, a proxy simulation env, and a frame from demos to overlay the sim observation (with robot arm removed); you can generate assets via https://www.sudo.ai/

MasterXiong commented 1 week ago

I see. Thanks! I wanted to run demos for existing simpler environments. I'll try with the first approach you suggest. Thanks!

MasterXiong commented 1 week ago

Hi @xuanlinli17 , thanks a lot for your help!

About "adjust the initial object pose to match the demo's first frame" in your previous reply, I was wondering that how can I achieve this adjustment? I checked the OXE data and there seems to be no information about the demo's initial pose. Or should I do this manually by checking their visual appearance by myself? Could you help provide a simple example code of how to do the adjustment here? Thanks!

xuanlinli17 commented 1 week ago

You have to manually tune the object pose to match demo's first frame. You can use the nvdiffrast utilities in https://github.com/Jiayuan-Gu/GeTex (except baking texture).