v0.6 (#18)

* Sampling over both env and denoising steps in DPPO updates (#13)

* sample one from each chain

* full random sampling

* Add Proficient Human (PH) Configs and Pipeline (#16)

* fix missing cfg

* add ph config

* fix how terminated flags are added to buffer in ibrl

* add ph config

* offline calql for 1M gradient updates

* bug fix: number of calql online gradient steps is the number of new transitions collected

* add sample config for DPPO with ta=1

* Sampling over both env and denoising steps in DPPO updates (#13)

* sample one from each chain

* full random sampling

* fix diffusion loss when predicting initial noise

* fix dppo inds

* fix typo

* remove print statement

---------

Co-authored-by: Justin M. Lidard <jlidard@neuronic.cs.princeton.edu>
Co-authored-by: allenzren <allen.ren@princeton.edu>

* update robomimic configs

* better calql formulation

* optimize calql and ibrl training

* optimize data transfer in ppo agents

* add kitchen configs

* re-organize config folders, rerun calql and rlpd

* add scratch gym locomotion configs

* add kitchen installation dependencies

* use truncated for termination in furniture env

* update furniture and gym configs

* update README and dependencies with kitchen

* add url for new data and checkpoints

* update demo RL configs

* update batch sizes for furniture unet configs

* raise error about dropout in residual mlp

* fix observation bug in bc loss

---------

Co-authored-by: Justin Lidard <60638575+jlidard@users.noreply.github.com>
Co-authored-by: Justin M. Lidard <jlidard@neuronic.cs.princeton.edu>

env/gym_utils/wrapper/furniture.py

@@ -132,9 +132,10 @@ class FurnitureRLSimEnvMultiStepWrapper(gym.Wrapper):
 
         nobs: np.ndarray = self.process_obs(obs)
         truncated: np.ndarray = truncated.squeeze().cpu().numpy()
-        terminated: np.ndarray = np.zeros_like(truncated, dtype=bool)
+        # terminated: np.ndarray = np.zeros_like(truncated, dtype=bool)
 
-        return {"state": nobs}, reward, terminated, truncated, info
+        # since we only assign reward at the timestep where one stage is finished, and reward does not accumulate, we consider the final step of the episode as terminal
+        return {"state": nobs}, reward, truncated, truncated, info
 
     def _inner_step(self, action_chunk: torch.Tensor):
         dense_reward = torch.zeros(action_chunk.shape[0], device=action_chunk.device)