long-term prediction - Githubissues

jannerm / trajectory-transformer

Code for the paper "Offline Reinforcement Learning as One Big Sequence Modeling Problem"

https://trajectory-transformer.github.io

MIT License

464 stars 65 forks source link

long-term prediction #17

Open qsa-fox opened 1 year ago

qsa-fox commented 1 year ago

Hi, how can I reproduce the Trajectory predictions in Figure 2 in the paper? Does this long-term prediction also need discrete states and actions? Hope for more details about the long-term prediction experiment implementations.

NiclasBottenbruchTum commented 2 months ago

Yes, it also requires discretization since the trajectory transformer only operates on discretized data. Appendix E Beam Search Hyperparameters in the paper seems to only specify the parameters for planning.

It would be really helpful if the authors @jannerm could state the hyperparameters for this rollout generation. Further questions are:

What is the used rollout policy for the PETS feedforward gaussian dynamics model?
Which dataset was used for the humanoid environment?

Best, Niclas