Project import generated by Copybara.

PiperOrigin-RevId: 283710365

gfootball/env/__init__.py

@@ -125,7 +125,7 @@ def create_environment(env_name='',
         composed of 4 planes of size 'channel_dimensions'.
         Its size is then 'channel_dimensions'x4 (or 'channel_dimensions'x16 when
         stacked is True).
-        The first plane P holds the position of players on the left team
+        The first plane P holds the position of players on the left
         team, P[y,x] is 255 if there is a player at position (x,y), otherwise,
         its value is 0.
         The second plane holds in the same way the position of players

gfootball/env/config.py

@@ -144,10 +144,10 @@ def update(self, config):
   def ScenarioConfig(self):
     return self._scenario_cfg
 
-  def NewScenario(self):
+  def NewScenario(self, inc = 1):
     if 'episode_number' not in self._values:
       self._values['episode_number'] = 0
-    self._values['episode_number'] += 1
+    self._values['episode_number'] += inc
     self._scenario_values = {}
     from gfootball.env import scenario_builder
     self._scenario_cfg = scenario_builder.Scenario(self).ScenarioConfig()

gfootball/env/football_env_core.py

@@ -52,15 +52,18 @@ class FootballEnvCore(object):
   def __init__(self, config):
     global _unused_engines
     self._config = config
+    self._sticky_actions = football_action_set.get_sticky_actions(config)
+    self._use_rendering_engine = False
     if _unused_engines:
       self._env = _unused_engines.pop()
     else:
       self._env = libgame.GameEnv()
     self._env.game_config.physics_steps_per_frame = config['physics_steps_per_frame']
-    self._sticky_actions = football_action_set.get_sticky_actions(config)
-    self._use_rendering_engine = False
+    # Reset is needed here to make sure render() API call before reset() API
+    # call works fine (get/setState makes sure env. config is the same).
+    self.reset(inc=0)
 
-  def _reset(self, animations):
+  def _reset(self, animations, inc):
     global _unused_engines
     global _unused_rendering_engine
     assert (self._env.state == GameState.game_created or
@@ -70,21 +73,21 @@ def _reset(self, animations):
     self._step = 0
     self._observation = None
     self._info = None
-    self._config.NewScenario()
+    self._config.NewScenario(inc=inc)
     if self._env.state == GameState.game_created:
       self._env.start_game()
     self._env.reset(self._config.ScenarioConfig(), animations)
     self._env.state = GameState.game_running
 
-  def reset(self):
+  def reset(self, inc=1):
     """Reset environment for a new episode using a given config."""
     self._episode_start = timeit.default_timer()
     self._action_set = football_action_set.get_action_set(self._config)
     trace = observation_processor.ObservationProcessor(self._config)
     self._cumulative_reward = 0
     self._step_count = 0
     self._trace = trace
-    self._reset(self._env.game_config.render)
+    self._reset(self._env.game_config.render, inc=inc)
     while not self._retrieve_observation():
       self._env.step()
     return True
@@ -363,15 +366,15 @@ def render(self, mode):
     if not self._env.game_config.render:
       if not self._use_rendering_engine:
         if self._env.state != GameState.game_created:
-          state = self.get_state()
+          state = self.get_state("")
           self.close()
           if _unused_rendering_engine:
             self._env = _unused_rendering_engine
             _unused_rendering_engine = None
           else:
             self._env = libgame.GameEnv()
           self._rendering_in_use()
-          self._reset(False)
+          self._reset(animations=False, inc=0)
           self.set_state(state)
           # We call render twice, as the first call has bad camera position.
           self._env.render(False)

gfootball/env/football_env_test.py

@@ -125,7 +125,7 @@ def check_determinism(self, extensive=False):
     for episode in range(1 if extensive else 2):
       hash_value = compute_hash(env, actions, extensive)
       if extensive:
-        self.assertEqual(hash_value, 4203104251)
+        self.assertEqual(hash_value, 2258127135)
       elif episode % 2 == 0:
         self.assertEqual(hash_value, 716323440)
       else:
@@ -169,7 +169,7 @@ def test_render(self):
     for _ in range(10):
       o, _, _, _ = env.step(football_action_set.action_right)
       hash = observation_hash(o, hash)
-    self.assertEqual(hash, 307836701)
+    self.assertEqual(hash, 845594868)
     env.close()
 
   def test_dynamic_render(self):
@@ -292,17 +292,24 @@ def test_player_order_invariant(self):
   @parameterized.parameters(range(1))
   def test_setstate(self, seed):
     """Checks setState functionality."""
-    cfg = config.Config({
+    cfg1 = config.Config({
         'level': 'tests.symmetric',
-        'game_engine_random_seed': seed
+        'game_engine_random_seed': seed,
+        'reverse_team_processing' : False
     })
-    env1 = football_env.FootballEnv(cfg)
-    env2 = football_env.FootballEnv(cfg)
+    cfg2 = config.Config({
+        'level': 'tests.symmetric',
+        'game_engine_random_seed': seed + 10,
+        'reverse_team_processing' : False
+    })
+    env1 = football_env.FootballEnv(cfg1)
+    env2 = football_env.FootballEnv(cfg2)
     initial_obs = env1.reset()
     env2.reset()
     initial_state = env1.get_state()
+    env2.set_state(initial_state)
     random.seed(seed)
-    actions = len(football_action_set.get_action_set(cfg))
+    actions = len(football_action_set.get_action_set(cfg1))
     first_action = random.randint(0, actions - 1)
     first_obs, _, _, _ = env1.step(first_action)
     _, _, _, _ = env2.step(first_action)

gfootball/env/remote_football_env.py

@@ -127,8 +127,10 @@ def _reset_with_retries(self, request):
         return stub.StartGame(request, timeout=10*60)
       except grpc.RpcError as e:
         if e.code() == grpc.StatusCode.DEADLINE_EXCEEDED:
+          time_to_sleep = 1
           continue
         logging.warning('Exception during request: %s', e)
+        logging.warning('Sleeping for %d seconds', time_to_sleep)
         time.sleep(time_to_sleep)
         if time_to_sleep < 1000:
           time_to_sleep *= 2

setup.py

@@ -41,7 +41,7 @@ def run(self):
 
 setup(
     name='gfootball',
-    version='2.0.2',
+    version='2.0.3',
     description=('Google Research Football - RL environment based on '
                  'open-source game Gameplay Football'),
     author='Google LLC',

third_party/gfootball_engine/src/cmake/backtrace.cpp

@@ -11,12 +11,23 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include <signal.h>
 #include <stdio.h>
-#include <stdlib.h>
+#include <execinfo.h>
+#include <signal.h>
 #include "backtrace.h"
+#include <stdlib.h>
+#include <unistd.h>
+
+void print_stacktrace() {
+  void *array[10];
+  size_t size;
+  size = backtrace(array, 10);
+  backtrace_symbols_fd(array, size, STDERR_FILENO);
+}
+
 
 void handler(int sig) {
+  print_stacktrace();
   printf("Error: signal %d:\n", sig);
   exit(1);
 }

third_party/gfootball_engine/src/cmake/backtrace.h

@@ -14,7 +14,7 @@
 #ifndef _CMAKE_BACKTRACE_H_
 #define _CMAKE_BACKTRACE_H_
 
-static inline void print_stacktrace() { }
+void print_stacktrace();
 
 void install_stacktrace();
 

third_party/gfootball_engine/src/game_env.cpp

@@ -96,6 +96,7 @@ void GameEnv::setConfig(ScenarioConfig& scenario_config) {
 
 void GameEnv::start_game() {
   assert(context == nullptr);
+  install_stacktrace();
   std::cout.precision(17);
   context = new GameContext();
   ContextHolder c(this);
@@ -348,8 +349,8 @@ void GameEnv::step() {
 void GameEnv::ProcessState(EnvState* state) {
   state->process(&this->state, sizeof(this->state));
   state->process(waiting_for_game_count);
-  context->gameTask->GetMatch()->ProcessState(state);
   context->ProcessState(state);
+  context->gameTask->GetMatch()->ProcessState(state);
 }
 
 void GameEnv::render(bool swap_buffer) {

third_party/gfootball_engine/src/main.cpp

@@ -200,7 +200,17 @@ void Tracker::verify_snapshot(long pos, int line, const char* file,
 
 void GameContext::ProcessState(EnvState* state) {
   state->process((void*)&rng, sizeof(rng));
-//  scenario_config->ProcessState(state);
+  if (state->Load()) {
+    EnvState reader(game, "");
+    game->scenario_config.ProcessStateConstant(&reader);
+    if (reader.GetState() != state->GetState().substr(state->getpos(),
+        reader.GetState().length())) {
+      Log(e_FatalError, "football", "set_state",
+          "Current environment scenario != scenario in the state.");
+    }
+  }
+  game->scenario_config.ProcessStateConstant(state);
+  game->scenario_config.ProcessState(state);
 #ifdef FULL_VALIDATION
   anims->ProcessState(state);
 #endif

third_party/gfootball_engine/src/main.hpp

@@ -133,7 +133,7 @@ struct ScenarioConfig {
     }
     return leftDistance < rightDistance;
   }
-  void ProcessState(EnvState* state) {
+  void ProcessStateConstant(EnvState* state) {
     state->process(ball_position);
     int size = left_team.size();
     state->process(size);
@@ -151,18 +151,20 @@ struct ScenarioConfig {
     state->process(right_agents);
     state->process(use_magnet);
     state->process(offsides);
-    state->process(real_time);
-    state->process(game_engine_random_seed);
-    state->process(reverse_team_processing);
     state->process(left_team_difficulty);
     state->process(right_team_difficulty);
     state->process(deterministic);
     state->process(end_episode_on_score);
     state->process(end_episode_on_possession_change);
     state->process(end_episode_on_out_of_play);
     state->process(game_duration);
+  }
 
- }
+  void ProcessState(EnvState* state) {
+    state->process(real_time);
+    state->process(game_engine_random_seed);
+    state->process(reverse_team_processing);
+  }
 };
 
 enum GameState {

third_party/gfootball_engine/src/onthepitch/match.cpp

@@ -653,6 +653,14 @@ void Match::UpdateIngameCamera() {
 }
 
 void Match::ProcessState(EnvState* state) {
+  if (state->getConfig()->reverse_team_processing) {
+    std::swap(first_team, second_team);
+  }
+  state->process(first_team);
+  state->process(second_team);
+  if (state->getConfig()->reverse_team_processing) {
+    std::swap(first_team, second_team);
+  }
   bool team_0_mirror = teams[0]->isMirrored();
   bool team_1_mirror = teams[1]->isMirrored();
   bool ball_mirror =

third_party/gfootball_engine/src/onthepitch/match.hpp

@@ -169,8 +169,8 @@ class Match {
 
     MatchData *matchData;
     Team *teams[2];
-    const int first_team = 0;
-    const int second_team = 1;
+    int first_team = 0;
+    int second_team = 1;
     bool ball_mirrored = false;
 
     Officials *officials;

third_party/gfootball_engine/src/onthepitch/teamAIcontroller.cpp

@@ -145,16 +145,9 @@ void TeamAIController::Process() {
   if (lineX * team->GetDynamicSide() - allowSlackDistance >
       deepestDanger * team->GetDynamicSide()) {
     DO_VALIDATION;
-    //printf("previous: %f\n", deepestDanger);
     deepestDanger = lineX - allowSlackDistance * team->GetDynamicSide();
-    //printf("now:      %f\n", deepestDanger);
   }
 
-  /*
-  if (team->GetID() == 0) SetRedDebugPilon(Vector3(deepestDanger, 0, 0));
-  if (team->GetID() == 1) SetYellowDebugPilon(Vector3(deepestDanger, 0, 0));
-  */
-
   offsideTrapX = deepestDanger;
   // disable
   //offsideTrapX = pitchHalfW * team->GetSide();
@@ -424,9 +417,9 @@ Vector3 TeamAIController::GetAdaptedFormationPosition(
   // maybe setting offside trap this way doesn't work too well: after all, defensive positioning stuff is done after this, thereby diminishing trap sometimes.
   // on the other hand, maybe this is a good basic trap setup, and we could apply the trap with the applyoffsidetrap function after the defensive positioning as well. (done from the def/mid strategies code)
   if (backXBound * team->GetDynamicSide() >
-      GetOffsideTrapX() * team->GetDynamicSide())
-    backXBound = GetOffsideTrapX();
-
+      GetOffsideTrapX() * team->GetDynamicSide()) {
+    backXBound = clamp(GetOffsideTrapX(), -pitchHalfW, pitchHalfW);
+  }
   float xFocus = 0.0f;
   float xFocusStrength = 0.0f;
   float yFocus = ballY * 1.0f;
@@ -1307,8 +1300,7 @@ void TeamAIController::ProcessState(EnvState *state) {
   DO_VALIDATION;
   state->process(taker);
   state->process(static_cast<void*>(&setPieceType), sizeof(setPieceType));
-  //baseTeamTactics.ProcessState(state);
-  //teamTacticsModMultipliers.ProcessState(state);
+  baseTeamTactics.ProcessState(state);
   liveTeamTactics.ProcessState(state);
   state->process(offensivenessBias);
   state->process(teamHasPossession);