Remove unused functions from mettagrid and grid_env (Metta-AI#274)

Cleans up mettagrid and grid_env by removing unused functions. This includes removing last_action tracking, which currently doesn't work. This cleaning is nice on it's own, but is mostly aimed towards making it easier to migrate this code to cpp.

Author: sasmith

configs/agent/reference_design.yaml

@@ -9,7 +9,6 @@ policy_selector:
 
 observations:
  obs_key: grid_obs
- track_last_action: false
 
 clip_range: 3.0 # clip weights greater or less than this value * largest weight in layer * clip_scale
 analyze_weights_interval: 300 # compute weight metrics every N epochs and log to wandb, add analyze_weights: true to the layer to enable

configs/agent/simple.yaml

@@ -10,7 +10,6 @@ policy_selector:
 
 observations:
  obs_key: grid_obs
- track_last_action: false
 
 clip_range: 0 # set to 0 to disable clipping
 analyze_weights_interval: 300

configs/env/mettagrid/mettagrid.yaml

@@ -3,7 +3,6 @@ _target_: mettagrid.mettagrid_env.MettaGridEnv
 
 report_stats_interval: 100
 normalize_rewards: false
-track_last_action: false
 
 sampling: 0
 

configs/env/mettagrid/puffer.yaml

@@ -3,7 +3,6 @@ _target_: mettagrid.mettagrid_env.MettaGridEnv
 
 report_stats_interval: 100
 normalize_rewards: false
-track_last_action: false
 
 sampling: 0
 

configs/user/sasmith.yaml

@@ -6,13 +6,9 @@ defaults:
  - _self_
 
 trainer:
- env: /env/mettagrid/navigation/training/multienv_with_cylinder
  evaluate_interval: 200
  evaluate: false
 
-sim:
- env: /env/mettagrid/navigation/training/multienv_with_cylinder
-
 eval_db_uri: wandb://artifacts/testing
 
 analyzer:
@@ -42,6 +38,6 @@ sweep:
  metric: action.use
 
 seed: null
-run_id: 20250429.2
+run_id: 20250501.1
 run: ${oc.env:USER}.local.${run_id}
 trained_policy_uri: ./train_dir/${oc.env:USER}.local.train.${run_id}/checkpoints

deps/mettagrid/README.md

@@ -94,10 +94,6 @@ The `use` action allows agents to interact with objects such as altars, converte
 
 The `swap` action allows agents to swap positions with other agents. It is currently not implemented.
 
-### Tracking the Agent's Last Action
-
-The `track_last_action` property adds the agent's last action to the observation space. It is currently not implemented.
-
 ## Configuration
 
 The MettaGrid environment is highly configurable through the use of YAML configuration files. These files specify the layout of the gridworld, the placement of objects, and various properties of the objects and agents.

deps/mettagrid/configs/benchmark.yaml

@@ -1,7 +1,6 @@
 # This is a modified `sample.yaml` config with sampling removed, for benchmark_env_perf.py script.
 report_stats_interval: 100
 normalize_rewards: false
-track_last_action: false
 
 sampling: 0
 

deps/mettagrid/configs/show_map.yaml

@@ -1,7 +1,6 @@
 # This config is used by `tools.map.view` script.
 report_stats_interval: 100
 normalize_rewards: false
-track_last_action: false
 
 game:
  # no map_builder, tools.utils.view passes the map directly

deps/mettagrid/configs/test_basic.yaml

@@ -1,7 +1,6 @@
 _target_: mettagrid.mettagrid_env.MettaGridEnv
 
 normalize_rewards: false
-track_last_action: false
 
 sampling: 0
 

deps/mettagrid/mettagrid/grid_env.pxd

@@ -31,8 +31,6 @@ cdef class GridEnv:
 
  unsigned short _obs_width
  unsigned short _obs_height
- unsigned short _middle_x
- unsigned short _middle_y
 
  vector[Agent*] _agents
 
@@ -86,7 +84,6 @@ cdef class GridEnv:
  cpdef list[str] grid_features(self)
  cpdef list[str] action_names(self)
  cpdef unsigned int num_agents(self)
- cpdef tuple observation_shape(self)
 
  cpdef tuple[cnp.ndarray, dict] reset(self)
  cpdef tuple[cnp.ndarray, cnp.ndarray, cnp.ndarray, cnp.ndarray, dict] step(self, cnp.ndarray actions)
@@ -109,6 +106,4 @@ cdef class GridEnv:
  cpdef dict get_episode_stats(self)
  cpdef get_episode_rewards(self)
 
- cpdef tuple get_buffers(self)
  cpdef cnp.ndarray render_ascii(self)
- cpdef cnp.ndarray grid_objects_types(self)