Initial commit

config/eval/cube.yaml

@@ -0,0 +1,61 @@
+defaults:
+  - launcher: local
+  - solver: cem
+  - _self_
+
+world:
+  env_name: swm/OGBCube-v0
+  num_envs: ${eval.num_eval}
+  max_episode_steps: ??? # make sure it's >= eval_budget
+  history_size: 1
+  frame_skip: 1
+  env_type: single
+  ob_type: states
+  multiview: False
+  width: 224
+  height: 224
+  visualize_info: False
+  terminate_at_goal: True
+
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+    - action
+
+seed: 42
+policy: random # ckpt name or random
+
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5 # frameskip
+
+# evaluation from dataset (replay expert trajectories)
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: ogbench/cube_single_expert
+  callables:
+    # -- set state
+    - method: set_state
+      args:
+        qpos:
+          value: qpos
+        qvel:
+          value: qvel
+    # -- set target pos
+    - method: set_target_pos
+      args:
+        cube_id:
+          value: 0
+          in_dataset: False
+        target_pos:
+          value: goal_privileged_block_0_pos
+        target_quat:
+          value: goal_privileged_block_0_quat
+
+output:
+  filename: ogb_cube_results.txt
+

config/eval/launcher/local.yaml

@@ -0,0 +1,7 @@
+# @package _global_
+# Local launcher configuration (no SLURM)
+
+defaults:
+  - override /hydra/launcher: basic
+
+cache_dir: null # use stable-worldmodel default cache

config/eval/pusht.yaml

@@ -0,0 +1,48 @@
+defaults:
+  - launcher: local
+  - solver: cem
+  - _self_
+
+world:
+  env_name: swm/PushT-v1
+  num_envs: ${eval.num_eval}
+  max_episode_steps: ??? # make sure it's >= eval_budget
+  history_size: 1
+  frame_skip: 1
+
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+    - action
+    - proprio
+    - state
+
+seed: 42
+policy: random # ckpt name or random
+
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5 # frameskip
+
+# evaluation from dataset (replay expert trajectories)
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: pusht_expert_train
+  callables:
+    # -- set state
+    - method: _set_state
+      args:
+        state:
+          value: state
+    # -- set goal state
+    - method: _set_goal_state
+      args:
+        goal_state:
+          value: goal_state
+          
+output:
+  filename: pusht_results.txt

config/eval/reacher.yaml

@@ -0,0 +1,50 @@
+defaults:
+  - launcher: local
+  - solver: cem
+  - _self_
+
+world:
+  env_name: swm/ReacherDMControl-v0
+  num_envs: ${eval.num_eval}
+  max_episode_steps: ??? # make sure it's >= eval_budget
+  history_size: 1
+  frame_skip: 1
+  task: qpos_match
+
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+    - action
+
+seed: 42
+policy: random # ckpt name or random
+
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5 # frameskip
+
+# evaluation from dataset (replay expert trajectories)
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: dmc/reacher_random
+  callables:
+    # -- set state
+    - method: set_state
+      args:
+        qpos:
+          value: qpos
+        qvel:
+          value: qvel
+
+    - method: set_target_qpos
+      args:
+        target_qpos:
+          value: goal_qpos
+
+output:
+  filename: dmc_results.txt
+

config/eval/solver/adam.yaml

@@ -0,0 +1,13 @@
+_target_: stable_worldmodel.solver.GradientSolver
+model: ???
+n_steps: 30
+batch_size: 1
+num_samples: 100
+action_noise: 0
+device: "cuda"
+seed: ${seed}
+optimizer_cls:
+  _target_: hydra.utils.get_class
+  path: torch.optim.AdamW
+optimizer_kwargs:
+  lr: 0.1

config/eval/solver/cem.yaml

@@ -0,0 +1,9 @@
+_target_: stable_worldmodel.solver.CEMSolver
+model: ???
+batch_size: 1
+num_samples: 300
+var_scale: 1.0
+n_steps: 30
+topk: 30
+device: "cuda"
+seed: ${seed}

config/eval/tworoom.yaml

@@ -0,0 +1,47 @@
+defaults:
+  - launcher: local
+  - solver: cem
+  - _self_
+
+world:
+  env_name: swm/TwoRoom-v1
+  num_envs: ${eval.num_eval}
+  max_episode_steps: ??? # make sure it's >= eval_budget
+  history_size: 1
+  frame_skip: 1
+
+seed: 42
+policy: random # ckpt name or random
+
+dataset:
+  stats: ${eval.dataset_name}
+  keys_to_cache:
+    - action
+    - proprio
+
+plan_config:
+  horizon: 5
+  receding_horizon: 5
+  action_block: 5 # frameskip
+
+# evaluation from dataset (replay expert trajectories)
+eval:
+  num_eval: 50
+  goal_offset_steps: 25
+  eval_budget: 50
+  img_size: 224
+  dataset_name: tworoom
+  callables:
+    # -- set state
+    - method: _set_state
+      args:
+        state:
+          value: proprio
+    # -- set goal state
+    - method: _set_goal_state
+      args:
+        goal_state:
+          value: goal_proprio
+
+output:
+  filename: tworoom_results.txt