[Fix] road_traffic map_type="2" crashes on GPU reset (#175)

* add tests failing for road_traffic map type 2

* fix failing tests for road_traffic map type 2

Author: vtalpaert

tests/test_scenarios/test_road_traffic.py

@@ -0,0 +1,69 @@
+#  Copyright (c) ProrokLab.
+#
+#  This source code is licensed under the license found in the
+#  LICENSE file in the root directory of this source tree.
+
+import pytest
+import torch
+
+from vmas import make_env
+
+
+class TestRoadTraffic:
+    def setup_env(self, n_envs, device="cpu", **kwargs) -> None:
+        self.env = make_env(
+            scenario="road_traffic",
+            num_envs=n_envs,
+            device=device,
+            continuous_actions=True,
+            **kwargs,
+        )
+        self.env.seed(0)
+
+    def _seed_buffer(self, device):
+        """Seed initial_state_buffer with a real state and force it to always be used."""
+        scenario = self.env.scenario
+        buf = scenario.initial_state_buffer
+        buf.add(scenario.state_buffer.get_latest(n=1)[0])
+        buf.probability_use_recording = torch.tensor(1.0, device=device)
+
+    @pytest.mark.parametrize("map_type", ["1", "2"])
+    def test_map_type_runs(self, map_type, n_envs=4, n_steps=10):
+        self.setup_env(n_envs=n_envs, map_type=map_type)
+        self.env.reset()
+        for _ in range(n_steps):
+            actions = [
+                torch.zeros(n_envs, agent.action.action_size)
+                for agent in self.env.agents
+            ]
+            obs, rews, dones, _ = self.env.step(actions)
+            if dones.any():
+                for env_index, done in enumerate(dones):
+                    if done:
+                        self.env.reset_at(env_index)
+
+    def test_map_type_2_reset_uses_buffer(self, n_envs=4):
+        self.setup_env(n_envs=n_envs, map_type="2")
+        self.env.reset()
+        actions = [
+            torch.zeros(n_envs, agent.action.action_size)
+            for agent in self.env.agents
+        ]
+        self.env.step(actions)
+        self._seed_buffer(device="cpu")
+        self.env.reset_at(0)
+
+    @pytest.mark.skipif(
+        not torch.cuda.is_available(),
+        reason="GPU required to reproduce road_traffic map_type=2 device bugs",
+    )
+    def test_gpu_map_type_2_rand_device(self, n_envs=4):
+        self.setup_env(n_envs=n_envs, device="cuda", map_type="2")
+        self.env.reset()
+        actions = [
+            torch.zeros(n_envs, agent.action.action_size, device="cuda")
+            for agent in self.env.agents
+        ]
+        self.env.step(actions)
+        self._seed_buffer(device="cuda")
+        self.env.reset_at(0)

vmas/scenarios/road_traffic.py

@@ -948,7 +948,8 @@ def reset_world_at(self, env_index: int = None, agent_index: int = None):
             if (
                 (self.parameters.map_type == "2")
                 and (
-                    torch.rand(1) < self.initial_state_buffer.probability_use_recording
+                    torch.rand(1, device=self.world.device)
+                    < self.initial_state_buffer.probability_use_recording
                 )
                 and (self.initial_state_buffer.valid_size >= 1)
             ):
@@ -1113,6 +1114,7 @@ def reset_init_state(
             agents[i_agent].set_pos(initial_state[i_agent, 0:2], batch_index=env_i)
             agents[i_agent].set_rot(initial_state[i_agent, 2], batch_index=env_i)
             agents[i_agent].set_vel(initial_state[i_agent, 3:5], batch_index=env_i)
+            return ref_path, path_id
         else:
             is_feasible_initial_position_found = False
             # Ramdomly generate initial states for each agent
@@ -2300,7 +2302,7 @@ def done(self):
         is_collision_with_lanelets = self.collisions.with_lanelets.any(dim=-1)
 
         if self.parameters.map_type == "2":  # Record into the initial state buffer
-            if torch.rand(1) > (
+            if torch.rand(1, device=self.world.device) > (
                 1 - self.initial_state_buffer.probability_record
             ):  # Only a certain probability to record
                 for env_collide in torch.where(is_collision_with_agents)[0]: