alpasim/src/wizard/configs/base_config.yaml at main · NVlabs/alpasim

353 lines (307 loc) · 13.4 KB
# \/ defines order of inheritance of configs in hydra. `config_schema` is defined in python code (not .yaml)
# and serves as the base skeleton. `_self_` means this file is applied on top of that schema.
  - config_schema
  - deploy: null # Must be overridden (e.g., deploy=local)
  - driver: null # Must be overridden (e.g., driver=vavam, driver=alpamayo1)
  - topology: null # Must be overridden (e.g., topology=1gpu)
  - trafficsim: null # Optional override
  - controller: null # Optional override
  - sensorsim: null # Optional override
  - optional scenes_catalog: internal # auto-override when internal plugin installed
  - optional image_defaults: internal # auto-override when internal plugin installed
# \/ hydra-specific configuration of legacy behavior (off)
    chdir: false
# This dictionary serves only to be references in other parts of the config file
  filesystem: ???
  driver_model: ???
  mpc_implementation: linear # "linear" (OSQP, faster) or "nonlinear" (CasADi)
  image_registry: ""
  base_image: "${defines.image_registry}alpasim-base:${repo-version:}"
  hf_cache: "${oc.env:HF_HOME,${oc.env:HOME}/.cache/huggingface}"
  # defaults will work out of the box \/
  drivers: "${defines.filesystem}/drivers"
  sensordata: "${defines.filesystem}/nre-artifacts"
  trafficsim_map_cache: "${defines.filesystem}/trafficsim/unified_data_cache"
  sensorsim_entrypoint: "/app/internal/scripts/pycena/runtime/pycena_run"
  helper: scripts
  vscode: sources/remote-vscode-server
  physics_cache_size: 16 # should match or exceed concurrent scenes to avoid cache thrashing
# \/ the wizard.* section defines properties global to all alpasim modules
  # Name of the run, used to identify the run in the databases. If null, will
  # use the SLURM job name if set.
  run_name: null
  # change to `SLURM` for ORD deployments, `DOCKER_COMPOSE` for local runs via
  # docker-compose and `DOCKER` for local runs via "raw" docker or `NONE` for
  # only generating the configs.
  run_method: "DOCKER_COMPOSE"
  # directory path to log the results of this run. ??? indicates it is mandatory to override this via cmdline args
  log_dir: ???
  # with wizard.dry_run=true we wouldn't start any containers and just print out the commands that would be executed.
  # note this performs various checks interanlly and thus may still not work if you're not on an ORD node
  dry_run: false
  # assert that all defined volume mounts exist at the start of the simulation
  validate_mount_points: true
  # the directory which will be searched for `.sqsh` files with images needed for simulation when running on ORD with
  # `wizard.run_method=SLURM`. The default location should contain all files required for running alpasim versions from
  # the repo but you may need to add your own cache if you're playing with custom-built images.
  sqshcaches:
    - "${defines.filesystem}/sqsh"
    - "${defines.filesystem}/sqsh/dev"
  # the different containers will use ports {baseport}, {baseport+1}, ...
  # usually not needed to modify
  baseport: 6000
  # Does not need to be touched. For slurm runs (wizard.run_method=SLURM) this will be replaced with the slurm job ID
  slurm_job_id: null
  # Services to run for simulation (e.g., ["driver", "sensorsim", "physics", "trafficsim", "controller", "runtime"])
  run_sim_services:
    ["driver", "sensorsim", "physics", "trafficsim", "controller", "runtime"]
  # TODO: undocumented options
  run_mode: "BATCH"
  timeout: 600
  latest_symlink: false
  # If set, the wizard will pull the driver code from the specified hash into
  # `${wizard.log_dir}/driver_code`. Can be useful for mounting into the
  # driver container for debugging.
  driver_code_hash: null
  # Used if `driver_code_hash` is set. Requires configured ssh keys for
  # pulling from gitlab, but can also point towards a local repo!
  driver_code_repo: null
  submitter: "unspecified" # An optional specifier of the submitter (used for tracking purposes)
  description: "unspecified" # An optional description of the run (used for tracking purposes)
  # Global log level for all alpasim services (DEBUG, INFO, WARNING, ERROR)
  log_level: INFO
  # Selection method (set one of these two)
  scene_ids:
    - clipgt-01d503d4-449b-46fc-8d78-9085e70d3554 # default OSS scene from huggingface (26.02 dataset)
  test_suite_id: null # Test suite ID to use (from suites_csv)
  # Limit the number of scenes to run (0 or negative means no limit)
  limit_to_first_n: 0
  scene_cache: "${defines.filesystem}/nre-artifacts"
  scenes_csv:
    - "${repo-relative:'data/scenes/sim_scenes.csv'}"
  suites_csv:
    - "${repo-relative:'data/scenes/sim_suites.csv'}"
# \/ services.* defines the individual components of the simulation. Each of them is deployed from an image so the layout of
# each item is similar. For the services that pull in local code (e.g., controller, runtime, eval) we mount the repo-relative
# `src/` directory into `/mnt/src` in the container and run from there. The virtual environment is reused from the container.
  # \/ sensor simulator for now is just NRE. We'll use it as an example on the structure of a container definition
  sensorsim:
    image: nvcr.io/nvidia/nre/nre-ga:26.02
    external_image: true
    # \/ volumes lets us mount host (ORD/local) containers to the running container
    volumes:
      - "${scenes.scene_cache}:/mnt/nre-data"
      - "${defines.sensordata}/ego-hoods:/mnt/ego-hoods"
    # \/ environments lets you set environment variables inside the container
    environments:
      # this may not be necessary but at least on COLOSSUS by default pytorch had really stupid
      # configuration running to excessively parallelizing everything
      - OMP_NUM_THREADS=1
    # \/ command is like docker entrypoint + command combined
    command:
      - "${defines.sensorsim_entrypoint}"
      - "serve-grpc"
      - "--port={port}" # {port} is a wizard-generated unique variable which enumerates {baseport}, {baseport+1}, ...
      - "--host=0.0.0.0" # the default container IP
      - "--artifact-glob=/mnt/nre-data/{sceneset}/**/*.usdz"
      - "--egocar-hood-dir=/mnt/ego-hoods"
      - "--no-enable-nrend"
      - "--download-cache-dir /tmp/nre-cache-dir" # unused
      - "--cache-size=${defines.nre_cache_size}" # as a rule of thumb n_concurrent_rollouts + 1 allows to avoid premature evictions
      - "--max-workers=4"
      - "--enable-editing-actors"
      # - "--enable-timing" # uncomment to enable timing information on the sensor simulator side
  # \/ services.drive is the driving policy (Alpamayo/HydraMDP, etc)
    image: ${defines.base_image}
    volumes:
      - "${defines.drivers}:/mnt/drivers"
      - "${wizard.log_dir}:/mnt/output" # optional output like the debug images
      - "${repo-relative:'src'}:/repo/src"
      - "${repo-relative:'plugins'}:/repo/plugins"
      - "${defines.hf_cache}:/root/.cache/huggingface"
    command:
      - "uv run -m alpasim_driver.main"
      - "--config-path=/mnt/output"
      - "--config-name=driver-config.yaml"
      - "host=0.0.0.0"
      - "port={port}"
    image: ${defines.base_image}
    volumes:
      - "${scenes.scene_cache}:/mnt/nre-data"
      - "${repo-relative:'src'}:/repo/src"
      - "${repo-relative:'plugins'}:/repo/plugins"
    environments:
      - WARP_CACHE_PATH='/mnt/warp'
    command:
      - "uv run physics_server"
      - "--host=0.0.0.0" # again, default container IP
      - "--port={port}" # arbitrary, needs to be updated in runtime-config.yaml
      - "--artifact-glob=/mnt/nre-data/{sceneset}/**/*.usdz"
      - "--use-ground-mesh=true"
      - "--cache-size=${defines.physics_cache_size}" # should match or exceed concurrent scenes to avoid cache thrashing
      - "--log-level=${wizard.log_level}"
  trafficsim:
    image: ${defines.base_image}
    command:
      - "echo 'not yet included'"
  controller:
    image: ${defines.base_image}
    volumes:
      - "${wizard.log_dir}/controller:/mnt/output"
      - "${repo-relative:'src'}:/repo/src"
      - "${repo-relative:'plugins'}:/repo/plugins"
    command:
      - "uv run python -m alpasim_controller.server"
      - "--port={port}"
      - "--log_dir=/mnt/output"
      - "--log-level=${wizard.log_level}"
      - "--mpc-implementation=${defines.mpc_implementation}"
  # \/ the orchestrating container - only 1 needed irrespectively of however many other there are
    image: ${defines.base_image}
    # \/ only starts if all the dependencies started correctly
    depends_on:
      - driver
      - sensorsim
      - physics
      - trafficsim
      - controller
    volumes:
      - "${scenes.scene_cache}:/mnt/nre-data"
      - "${wizard.log_dir}:/mnt/log_dir"
      - "${or:${wizard.array_job_dir},${wizard.log_dir}}:/mnt/array_job_dir"
      - "${repo-relative:'src'}:/repo/src"
      - "${repo-relative:'plugins'}:/repo/plugins"
    gpus: null # uses no GPUs
    command:
      - "uv run python -m alpasim_runtime.simulate"
      - "--usdz-glob=/mnt/nre-data/{sceneset}/**/*.usdz"
      - "--user-config=/mnt/log_dir/{runtime_config_name}"
      - "--network-config=/mnt/log_dir/generated-network-config.yaml"
      - "--log-dir=/mnt/log_dir"
      - "--log-level=${wizard.log_level}"
      - "--array-job-dir=/mnt/array_job_dir"
      - "--eval-config=/mnt/log_dir/eval-config.yaml"
    replicas_per_container: 1 # should only be one runtime orchestrator
  # nr_workers and endpoints.*.n_concurrent_rollouts are set by topology configs.
  endpoints:
    # shut down the system after simulation is finished. without this flag the microservice servers
    # will remain on forever requring a manual interrupt (useful for debugging)
    do_shutdown: true
    # Cache size used by the scene cache monitor to ensure enough capacity
    sensorsim_cache_size: ${defines.nre_cache_size}
  enable_autoresume: false
  # How many scenes (in particular maps) to cache in the worker local artifact cache.
  artifact_cache_size: 10
  simulation_config:
    n_sim_steps: 200 # how many steps to simulate in a rollout
    n_rollouts: 1 # how many rollouts to simulate for that scenario
    force_gt_duration_us: 1_700_000 # 1.7s
    send_recording_ground_truth: false # Disable sending ground truth data to the driver
    control_timestep_us: 100_000
    pose_reporting_interval_us: 0 # 0 = controller reports only the final pose
    time_start_offset_us: 300_000
    ego_mask_rig_config_id: "hyperion_8_1" # which rig/directory in ego-hoods to use for ego masking
    planner_delay_us: 0
    # Actors present shorter than this (in micro-seconds) are ignored. Set to 0 to disable.
    min_traffic_duration_us: 3_000_000
    assert_zero_decision_delay: true
    physics_update_mode: "EGO_ONLY"
    route_generator_type: "MAP" # "MAP" or "RECORDED"
    vehicle: null # use values from the .usdz file (original dimensions)
    # Note: If you change the `frame_interval_us`, it will affect the frequency of
    # camera images sent to the driver, which, depending on the driver, could effect
    # closed loop behavior. If you want to keep the behavior unchanged, consider
    # asjusting the `driver.inference.Cframes_subsample` parameter,
    # which controls how many frames are skipped between consecutive images sent to the driver.
    cameras:
      - height: 320
        width: 512
        logical_id: camera_front_wide_120fov
        # If you change this see note above about `driver.inference.Cframes_subsample`
        frame_interval_us: 100_000
        shutter_duration_us: 30_000
        first_frame_offset_us: -30_000
      - height: 320
        width: 512
        logical_id: camera_front_tele_30fov
        # If you change this see note above about `driver.inference.Cframes_subsample`
        frame_interval_us: 100_000
        shutter_duration_us: 30_000
        first_frame_offset_us: -30_000
  # Set to false to disable in-runtime evaluation (e.g., for debugging or performance testing)
  enabled: true
    incl_road_edges: true
    incl_traffic_signs: true
    incl_wait_lines: true
    max_num_lanes: 20
    num_pts_per_lane: 20
  num_processes: 16
    vehicle_corner_roundness: 0.5
    vehicle_shrink_factor: 0.02
    min_ade:
      time_deltas: [0.5, 1.0, 2.5, 5.0]
      incl_z: False
      target: GT
    plan_deviation:
      incl_z: False
      avg_decay_rate: 0.1
      min_timesteps: 5
      camera_logical_id: camera_front_wide_120fov
  aggregation_modifiers:
    max_dist_to_gt_trajectory: 4.0
    render_video: True
    video_layouts: ["DEFAULT"] # ["DEFAULT", "REASONING_OVERLAY"]
    camera_id_to_render: camera_front_wide_120fov
    reasoning_text_refresh_interval_s: 1.0 # for reasoning overlay layout only
    overlay_plans_on_camera: True
    render_every_nth_frame: 1
    generate_combined_video: False
    combined_video_speed_factor: 0.33
    metrics_table_entries:
      - offroad_or_collision_at_fault
      - collision_any
      - collision_at_fault
      - collision_front
      - collision_lateral
      - collision_rear
      - offroad
      - dist_to_gt_trajectory
      - dist_to_gt_location
      - progress
      - progress_rel
      - safety_monitor_triggered
    map_video:
      map_radius_m: 20
      ego_loc: BOTTOM_CENTER
      rotate_map_to_ego: True
      map_elements_to_plot:
        - ROAD_LANE_CENTER
        - ROAD_LANE_LEFT_EDGE
        - ROAD_LANE_RIGHT_EDGE
        - ROAD_EDGE
        - STOP_LINE
        - GT_LINESTRING
        - EGO_GT_GHOST_POLYGON
        - DRIVER_RESPONSES
        - ROUTE
        - AGENTS
# driver config is loaded from the driver group (e.g., driver=vavam)
Provide feedback

Saved searches

Use saved searches to filter your results more quickly

FilesExpand file tree

base_config.yaml

Latest commit

History

base_config.yaml

File metadata and controls
