| import madrona_gpudrive |
| import pytest |
| import torch |
|
|
| @pytest.fixture(scope="module") |
| def sim_init(): |
| reward_params = madrona_gpudrive.RewardParams() |
| reward_params.rewardType = madrona_gpudrive.RewardType.DistanceBased |
| reward_params.distanceToGoalThreshold = 1.0 |
| reward_params.distanceToExpertThreshold = 1.0 |
|
|
| params = madrona_gpudrive.Parameters() |
| params.polylineReductionThreshold = 0.5 |
| params.observationRadius = 10.0 |
| params.collisionBehaviour = madrona_gpudrive.CollisionBehaviour.AgentStop |
| params.rewardParams = reward_params |
| params.maxNumControlledAgents = 2 |
| params.IgnoreNonVehicles = True |
| params.dynamicsModel = madrona_gpudrive.DynamicsModel.InvertibleBicycle |
| sim = madrona_gpudrive.SimManager( |
| exec_mode=madrona_gpudrive.madrona.ExecMode.CPU, |
| gpu_id=0, |
| scenes=["tests/pytest_data/test.json"], |
| params=params |
| ) |
|
|
| return sim |
|
|
| def test_forward_inverse_dynamics(sim_init): |
| sim = sim_init |
| |
| valid_agent_idx = 1 |
| done_tensor = sim.done_tensor().to_torch() |
| expert_trajectory_tensor = sim.expert_trajectory_tensor().to_torch() |
| action_tensor = sim.action_tensor().to_torch() |
| absolute_obs = sim.absolute_self_observation_tensor().to_torch() |
| self_obs = sim.self_observation_tensor().to_torch() |
| actions = torch.zeros_like(action_tensor) |
|
|
| traj = expert_trajectory_tensor[:,valid_agent_idx].squeeze() |
| pos = traj[:2*91].view(91,2) |
| vel = traj[2*91:4*91].view(91,2) |
| headings = traj[4*91:5*91].view(91,1) |
| invActions = traj[6*91:16*91].view(91,10) |
| print('invActions', invActions[:2]) |
| position = absolute_obs[0,valid_agent_idx,:2] |
| heading = absolute_obs[0,valid_agent_idx,7] |
| speed = self_obs[0, valid_agent_idx, 0] |
| assert torch.allclose(position, pos[0], atol=1e-2), f"Position mismatch: {position} vs {pos[0]}" |
| assert pytest.approx(heading.item(), abs=1e-2) == headings[0].item(), f"Heading mismatch: {heading.item()} vs {headings[0].item()}" |
| assert pytest.approx(speed.item(), abs=1e-2) == torch.norm(vel[0]).item(), f"Speed mismatch: {speed.item()} vs {torch.norm(vel[0]).item()}" |
|
|
| actions[:,valid_agent_idx,:3] = invActions[0,:3] |
| action_tensor.copy_(actions) |
| sim.step() |
| |
| assert torch.allclose(position, pos[1], atol=2e-2), f"Position mismatch: {position} vs {pos[1]} at step {1}" |
| assert pytest.approx(heading.item(), abs=3e-3) == headings[1].item(), f"Heading mismatch: {heading.item()} vs {headings[1].item()}" |
| assert pytest.approx(speed.item(), abs=1e-3) == torch.norm(vel[1]).item(), f"Speed mismatch: {speed.item()} vs {torch.norm(vel[1]).item()}" |
