This document lists major updates which change UX and require adaptation. It should be sorted by date (more recent on top) and link to MRs which introduce the changes.
This sync expands the public scene catalog, improves rollout reliability and scheduling, and adds new telemetry, evaluation, and Slurm controls.
The public 26.04 catalog is available as the public_2604 suite, containing
1,606 scenes after removing known-invalid artifacts. Scene-suite CSV rows now
pin an exact artifact with test_suite_id, scene_id, and uuid; suite lookup
no longer selects the newest artifact sharing a scene ID.
Artifacts whose parquet vector map is stored under fastmap/ are now
supported. AlpaSim loads it after the preferred map_data/ and clipgt/
sources, aligns its height to the reconstruction ground mesh, and retains XODR
as the final fallback.
Migration: Add a matching uuid column to custom suite CSVs. Every
(scene_id, uuid) pair must exist in the corresponding scene catalog.
Failed rollouts are retried up to runtime.max_rollout_retries times, with a
default of two retries. Renderer startup now waits for channel readiness and
retries known transient NRE initialization failures. Invalid vector maps return
the structured ROLLOUT_ERROR_CODE_INVALID_SCENE_MAP error code.
Scene-affine renderer dispatch now spreads non-affine work across the
least-loaded renderers and bounds scene replication. New controls include
max_renderers_per_scene, max_scenes_per_renderer, and cache refresh timing.
Migration: Replace a boolean override such as
runtime.scene_affine_dispatch=true with
runtime.scene_affine_dispatch.enabled=true. Public NRE releases without the
loaded-scene introspection RPC should leave this disabled.
Traffic simulation is explicitly disabled by default through
trafficsim=disabled; select trafficsim=catk to enable CATK. CATK now batches
requests, limits per-request CPU threading, and derives worker counts from the
selected topology.
Evaluation adds the open_loop_collision metric, which checks planned ego
trajectories against recorded actors over a configurable horizon. Force-GT
warmup exclusion is now anchored to the actual renderer handover, preventing
warmup frames from leaking into aggregated metrics. Route generation can be
disabled with runtime.simulation_config.route_generator_type=NONE.
The public Alpamayo 1 driver can derive deterministic per-inference seeds from
the session seed with driver.model.force_determinism=true. The low-speed MPC
linearization also corrects two erroneous derivative terms.
Migration: Custom CATK configs should use
catk.loader.prediction_steps instead of minimum_future_steps.
Per-process Prometheus exporters can now be scraped by an external Prometheus
without starting the wizard-managed sidecar. Set
wizard.prometheus.start_prometheus=false; discovery publication and exporter
lifecycle remain active. Slurm telemetry adds GPU ownership metrics for joining
AlpaSim processes with DCGM data, and the Grafana dashboard includes the new
scheduler, GPU, process, and throughput panels.
Slurm deployments can opt into CUDA MPS with wizard.enable_mps=true to allow
co-located GPU services to execute kernels concurrently. This option fails fast
for non-Slurm run methods. Slurm submission also supports automatic resubmission
of timed-out jobs.
AlpaSim runs now start Prometheus telemetry by default. The wizard allocates
metrics ports, starts Prometheus support services, writes scrape configuration,
and preserves the Prometheus TSDB under the run directory. Runs still generate
metrics_plot.png.
Grafana dashboard resources and a helper script are available for inspecting local or shared Prometheus file-SD targets:
src/tools/scripts/start-prometheus-grafana.sh <file-sd-dir>Eval aggregation no longer reads Prometheus runtime metric summaries or adds
runtime performance fields to driving metric outputs. Runs no longer generate
prometheus/runtime_metrics_summary.json; query Prometheus or use Grafana for
runtime performance analysis.
Added an integrated CATK-backed traffic simulation service. Use
trafficsim=catk to run the SMART/CATK traffic predictor from the AlpaSim base
image with model weights stored under data/trafficsim-models.
Bumped the base_config.yaml renderer image to nvcr.io/nvidia/nre/nre-ga:26.04 and repointed defines.renderer_entrypoint to the /app/run symlink (the old path was renamed in 26.04).
Eval aggregation now writes results-summary.json with per-rollout pass/fail status,
scene scores, failed rollout rows, and driver drive RPC latency from telemetry.
Ground-truth scoring adds progress_rel_to_total and gt_dist_traveled_m; tune or
disable scoring with eval.scene_score, and set eval.allow_aggregation_with_failed_rollouts=true
to aggregate successful rollouts while preserving failed rollout records.
This sync focuses on renderer unification, video-model support, scene catalog updates, and runtime robustness improvements.
Wizard-managed renderer services now use the renderer service key for both
the default NRE-backed renderer and video-model renderers. The runtime selects
the active renderer with runtime.renderer.kind, and generated configs now use
runtime.endpoints.renderer / network.renderer instead of sensorsim.
Video-model renderer support is now built into the main runtime and wizard
configuration. New public configs include deploy=external_video_model,
deploy=managed_flashdreams, +chunking=<8frame|12frame|16frame>, and
docs/VIDEO_MODEL.md. Docker Compose deployments now stop backing services when
runtime-0 exits.
Migration: Replace services.sensorsim and
wizard.run_sim_services=["sensorsim", ...] with services.renderer and
wizard.run_sim_services=["renderer", ...]. Replace
physics/implemented_in_sensorsim.yaml with
physics/implemented_in_renderer.yaml. Replace custom
runtime.endpoints.sensorsim / network.sensorsim entries with renderer.
For one-shot Compose runs, use docker compose up --exit-code-from runtime-0.
The default public Hugging Face scene catalog now uses the 26.01 release and the
public_2601 suite. The previous 25.05 catalog remains available through
sim_scenes_2505.csv / sim_suites_2505.csv.
Runtime scene discovery now uses scene_provider configuration instead of
passing artifact globs through the runtime CLI.
Migration: Existing 25.05 cached artifacts are not reused by the new default.
Pin scenes.scenes_csv / scenes.suites_csv to the legacy CSVs if you need the
old suite. For direct python -m alpasim_runtime.simulate usage, remove
--usdz-glob and configure scene_provider.kind plus
scene_provider.usdz.data_dir.
The runtime daemon now exposes RuntimeService.get_runtime_info for client-side
capacity and scene discovery. Startup logs now report connection progress while
waiting for services.
Closed-loop and RL workflows also gained DriveResponse.terminate_session,
single-rollout RolloutSpec.session_uuid, explicit driver precondition errors,
spawned evaluation workers, empty driver-response handling, and continuous
min-distance eval scorers.
Migration: Existing clients remain compatible. Daemon clients can call
get_runtime_info, set RolloutSpec.session_uuid for precise single-rollout
abort handling, or set DriveResponse.terminate_session=true for early episode
termination. Eval configs can opt into min_distance_to_obstacle_m and
min_distance_to_lane_boundary_m.
Rollout timing now starts from the first GT camera frame timestamp. Force-GT
startup blends recorded GT trajectories into physics-derived trajectories, and
runtime.simulation_config.skip_driver_during_force_gt can skip expensive driver
policy queries during force-GT warmup.
alpasim_utils.asl_to_frames now exports video-model RGB / HD-map streams and
names output frames by end-of-frame timestamp. alpasim_utils.print_asl redacts
large video-model image payloads.
Migration: Remove custom time_start_offset_us and camera
first_frame_offset_us settings. Update downstream scripts that parse frame
filenames to expect end-of-frame timestamps.
The src/grpc package now builds generated protobuf artifacts during package
builds, so downstream Git installs no longer need a manual compile-protos
step. trajdata-alpasim is pinned to a specific Git commit for consistent
workspace resolution.
Migration: Local proto development can still use uv run compile-protos.
Re-run dependency sync after pulling this change.
Wizard run modes now distinguish one-shot runtime execution from long-running runtime daemon deployment. wizard.run_mode=BATCH has been renamed to wizard.run_mode=ONESHOT, and wizard.run_mode=SERVER starts the runtime as a gRPC server for request-scoped simulations.
Migration: If you explicitly set wizard.run_mode=BATCH, replace it with wizard.run_mode=ONESHOT. The default behavior remains unchanged for standard one-shot execution workflows.
External driver ownership is now expressed through driver_source config groups rather than deployment targets:
- Default managed driver behavior remains unchanged.
driver_source=external_staticuses configured external driver addresses, replacing the removeddeploy=local_external_driverdeployment target.driver_source=external_dynamicsupports per-request drivers viaSimulationRequest.available_drivers.
Migration: Replace deploy=local_external_driver with deploy=local driver_source=external_static driver=manual.
Runtime evaluation now runs in its own thread instead of inline in the simulation loop. This decouples eval latency from the simulation step, improving throughput when evaluation is expensive.
Added torchmetrics>=1.8.2 to override-dependencies in the root pyproject.toml to resolve a conflict between upstream driver dependencies.
The Hydra config discovery plugin now detects YAML files that exist at the same relative path in multiple config providers (e.g. both wizard and an installed plugin). Duplicate paths raise a ValueError at startup, preventing silent config shadowing.
Driver config names, entry points, and model_type values now use explicit names instead of abbreviations:
| Before | After |
|---|---|
driver=ar1 |
driver=alpamayo1 |
driver=a15 |
driver=alpamayo1_5 |
Migration: Replace driver=ar1 with driver=alpamayo1 and driver=a15 with driver=alpamayo1_5 in CLI invocations, SLURM scripts, and any custom configs that reference these drivers.
The OSS sensorsim image has been upgraded from docker.io/carlasimulator/nvidia-nurec-grpc:0.2.0 to nvcr.io/nvidia/nre/nre-ga:26.02.
- The sensorsim entrypoint (
/app/run serve-grpc) and all shared flags are now defined once inbase_config.yaml. - New flag
--enable-editing-actorsadded to the base sensorsim command, required by NRE 26.3 for render requests that include dynamic object updates.
Migration: If you override services.sensorsim.command in a custom manifest, add --enable-editing-actors to the argument list.
The wizard config is now composed from three required, independent axes instead of monolithic deploy configs:
uv run alpasim_wizard deploy=local topology=1gpu driver=vavam wizard.log_dir=./out| Group | Purpose | Examples |
|---|---|---|
deploy= |
Where to run (filesystem, run method) | local, local_external_driver |
topology= |
GPU layout, replicas, concurrency | 1gpu, 2gpu, 8gpu_64rollouts |
driver= |
Which driving model | vavam, ar1, a15, manual |
All three are required. Omitting any prints a helpful error listing available options.
Each driver config now includes its own runtime settings via the Hydra defaults list. Specify a single config instead of a list:
| Before | After |
|---|---|
driver=[vavam,vavam_runtime_configs] |
driver=vavam |
driver=[ar1,alpamayo_runtime_configs] |
driver=ar1 |
driver=[a15,alpamayo_runtime_configs] |
driver=a15 |
The stable_manifest config group (oss.yaml, oss_gitlab.yaml) has been removed. Its content has been merged into base_config.yaml:
- Services built from the repo (driver, physics, controller, trafficsim, runtime) use
${defines.base_image}, which reads the version frompyproject.tomlat runtime via arepo-version:OmegaConf resolver. - The external sensorsim image (
nvcr.io/nvidia/nre/nre-ga:26.02) is set directly inbase_config.yaml. - A default OSS scene ID is now in
base_config.yaml, so new users can run without specifying scenes.
runtime.nr_workers and all runtime.endpoints.*.n_concurrent_rollouts values are now set by topology configs instead of base_config.yaml. Each topology preset defines capacity to match its GPU layout. base_config.yaml retains only behavioral settings (do_shutdown, enable_autoresume, etc.).
The scattered model/, experiment/, sim/, and exp/ config directories have been consolidated under a single exp/ group. Presets (e.g., vavam_4hz) moved to exp/presets/.
New optional groups in base_config.yaml defaults allow overriding service-specific settings:
controller=— override controller configsensorsim=— override NRE imagetrafficsim=— override trafficsim config
submit.shno longer defaults to any deploy target. All three axes (deploy=,topology=,driver=) must be specified.- Early sanity check rejects submissions with missing required config groups before allocating SLURM resources.
- Example:
sbatch submit.sh deploy=ord topology=8gpu_64rollouts driver=vavam
+deploy=syntax is nowdeploy=(no+prefix). Same fortopologyanddriver.driver=[<model>,<runtime_configs>]list syntax is now justdriver=<model>.cameras/wide_only_cam.yamlremoved (usecameras/1cam.yaml).stable_manifestconfig group removed entirely.- Deleted monolithic deploy configs:
iad_oss,ord_oss,ord_oss_single,local_2gpus,iad(OSS). Usedeploy=<target> topology=<layout>instead. runtime.nr_workersandruntime.endpoints.*defaults removed frombase_config.yaml(set by topology).defines.nre_cache_sizeremoved frombase_config.yaml(set by topology).
Alpamayo 1.5 is now available as a driver (model_type: a15). Use driver=a15 to run with the 10B model.
- New
A15Modeldriver with camera-index-aware inference and optional classifier-free guidance navigation (use_classifier_free_guidance_nav: true, ~60 GB VRAM). - AR1 and A1.5 now share a common
AlpamayoBaseModelbase class, reducing code duplication. planner_delay_usnow defaults to0everywhere; the legacyalpamayo_runtime_configsfile (which set 200ms delay) has been removed.
References to ~/.netrc in the Dockerfile and wizard's Docker Compose generation were unconditional, requiring all users to have the file. The Dockerfile now conditionally sets NETRC only when the secret is provided, and the wizard only includes the netrc secret in the compose config when ~/.netrc exists on the host.
The root pyproject.toml now exposes every workspace member as a named optional-dependency extra, enabling composable installs from the repo root. A bare uv sync installs nothing (avoiding heavy deps like torch by default).
uv sync --extra wizard— wizard and its transitive deps onlyuv sync --extra all— all core packagessource setup_local_env.shstill works and installs all core packages (plugins must be added separately).
See Onboarding — Dependency management for details.
Wizard config groups (e.g. driver, deploy) can now be extended by any installed package. Packages register an alpasim.configs entry point pointing to a Python package that contains YAML files, and the wizard automatically adds it to Hydra's search path at startup via SearchPathPlugin.
model_typein driver config is now a plain string (e.g."ar1","manual") instead of an enum.- The transfuser driver configs have been moved out of the wizard into the transfuser plugin — when installed,
driver=[transfuser,transfuser_runtime_configs]resolves automatically.
Alpasim is now extensible via Python entry points. Any installed package can register models, controllers, configs, or tools without modifying the core codebase.
- New
alpasim-pluginspackage (src/plugins) provides aPluginRegistrythat discovers entry points lazily at runtime. - Driver models (ar1, transfuser, vam, manual) and controller MPCs (linear, nonlinear) are registered as entry points and resolved by name.
- Run
uv run alpasim-infoto list all installed plugins.
See Plugin System for the full architecture, entry-point groups, and how to create new plugins.
- The runtime simulation loop is now event-based instead of a fixed sequential control-step loop.
pose_reporting_interval_usis the active pose-reporting setting; olderegopose_*configuration naming has been removed from the active runtime path.- The active egomotion noise model path was removed, so configs and tooling should no longer expect
egomotion_noisebehavior in standard runtime execution.
- The runtime can now run as a long-lived gRPC daemon that accepts simulation requests on demand.
- The gRPC API changed:
RolloutSpec.random_seedwas replaced bynr_rollouts, structured rollout results are returned, and ashut_downRPC was added for graceful shutdown. - One-shot CLI execution still works, but now routes through the same daemon engine internally.
- The manual scene artifact compatibility matrix was removed. Scene selection now treats newer NRE versions as backwards-compatible and chooses the newest available artifact per scene.
- Sensorsim/NRE scaling now relies on internal workers (
--max-workers) rather than multiple replicas per container in the common OSS deploy configs. - If you tune throughput, update your expectations for sensorsim capacity:
replicas_per_containeralone no longer tells the full story.
Added higher frequency pose/state information for when model updates are more sparse.
Additionally, changed the way that the HF_HOME environment variable is handled to be more like the public repo.
- ARM64 support: AlpaSim can now run on aarch64 (DGX Spark, DGX Station, IPP5 GB300).
Build with
docker build --secret id=netrc,src=$HOME/.netrc -t alpasim-base:arm64 .and deploy with+deploy=local_arm(Docker Compose) or+deploy=ipp5(SLURM). - Unified SLURM script:
src/tools/run-on-slurm/is the single entry point; previous per-site directories have been consolidated intosrc/tools/run-on-slurm/submit.sh.
Migration: Update SLURM submit commands:
cd src/tools/run-on-slurm && sbatch --account=<acct> --partition=<part> submit.sh +deploy=ord_osscd src/tools/run-on-ipp5 && sbatch submit.sh→cd src/tools/run-on-slurm && sbatch --account=<acct> --partition=<part> --gpus-per-node=4 submit.sh +deploy=ipp5
The wizard output directory structure has been reorganized for clarity:
./asl/directory renamed to./rollouts/- contains rollout logs organized by scene and session0.asland0.rclogfiles renamed torollout.aslandrollout.rclog./metrics/directory renamed to./telemetry/- contains Prometheus telemetry data (not to be confused with evaluation metrics stored in rollouts)- Videos are now saved next to ASL files:
rollouts/<scene_id>/<rollout_uuid>/<video>.mp4 - Metrics parquet files are saved next to ASL files:
rollouts/<scene_id>/<rollout_uuid>/metrics.parquet aggregate/videos/allnow uses symlinks instead of hard copies for space efficiency
Migration: If you have scripts that reference the old paths, update them to use the new structure:
asl/→rollouts/0.asl→rollout.asl0.rclog→rollout.rclogmetrics/→telemetry/eval/videos/orvideos/→rollouts/<scene_id>/<rollout_uuid>/<video>.mp4
- Evaluation metrics are now computed during simulation (in-runtime) by default, eliminating the need for separate eval containers.
- The previous behavior (running evaluation in separate containers after simulation) can be restored with
+eval=eval_in_separate_job. - This change simplifies the default workflow and reduces resource usage for most use cases.
- Videos are now saved next to ASL files in
rollouts/<scene_id>/<rollout_uuid>/(unified path for both modes). - TODO: Image-based metrics are not yet supported in this workflow (e.g. is_camera_black)
Removed maglev.av dependency from the base image to better align with the public-facing
repository. The dependency was required to produce roadcast logs, and this functionality has been
moved to a separate tool (along with the buildauth script) in src/tools/asl_to_roadcast. See the
README there for instructions on how to use it to generate roadcast logs going forward and how to
view the produced roadcast logs in DDB. Additionally, ddb and avmf have been removed since these
depended on having roadcast logs and weren't being used.
Added a new controller implementation in the OSS controller which is faster than the previous one
and allow the choice at runtime between the two implementations. The new (linear) implementation is
the default, and the nonlinear one can be selected using the defines.mpc_implementation wizard
configuration parameter.
Local directory support was recently dropped in one of the larger refactorings. This has been
restored with a slightly different interface. Now, for users to run Alpasim with local USDZ files,
they can use the scenes.local_usdz_dir configuration parameter. For example:
# to run all scenes in the local_usdz_dir directory:
alpasim_wizard +deploy=local wizard.log_dir=<output_dir> scenes.local_usdz_dir=<abs or rel path to directory> scenes.test_suite_id=local
# to run a subset of the scenes:
alpasim_wizard +deploy=local wizard.log_dir=<output_dir> scenes.local_usdz_dir=<abs or rel path to directory> scenes.scene_ids=[<your scene ids>]- A helper script
src/tools/run-on-slurm/resume_slurm_job.shis provided to simplify resuming failed array job tasks.
- Adds the ability for users to restart failed jobs in a batch by setting
runtime.enable_autoresume=true.
Three alpasim repositories are deprecated in favor of this one, to unify the development process more.
Using uv allows us to automatically updated wizard dependencies without future
action from the user, while currently users have to re-install the wizard.
To migrate:
- Install
uvif not yet done:curl -LsSf https://astral.sh/uv/install.sh | shAlternatively, runuv self updateas older versions have been reported to not work. - Install wizard:
uv tool install -e src/wizard/ condais no longer used with Alpasim, thealpasimenv can be deleted.
For developers only: For debugging using vscode I did the following:
uv syncinsrc/wizard/creates a venv undersrc/wizard/.venv- In
launch.json, use"module": "alpasim_wizard" - Use the command "Python: Select Interpreter" to manually pick the python
interpreter unter
.venv(you might need to enter the path as the venv wasn't picked up automatically for me).
- User facing:
runtime.endpoints.*.n_concurrent_batchesis now calledruntime.endpoints.*.n_concurrent_rollouts.runtime.batch_sizeno longer exists.
- Developer:
- The concept of batch size has been removed from the runtime.
- Instead of
Bound/UnboundBatchandRolloutwe haveBound/UnboundRollout.
- Instead of
- gRPC API changes.
- Fields like
batch_sizecan be assumed to always be equal to 1 androllout_indexequal to 0. They are deprecated. - Fields which are
repeatedto support multiple rollouts are deprecated. New fields (with single rollout per message semantics) are added. - Runtime falls back to deprecated fields - no breaking change for now.
- Fields like
- The concept of batch size has been removed from the runtime.
- Scene selection is performed via
scenes.{scene_ids,test_suite_id}instead ofwizard.nre_sceneset.- The options are mutually exclusive.
- Specific artifacts will be automatically selected to match the configured NRE version. If impossible, an error is thrown.
usdzfiles are now cached by theiruuidrather than path.python -m alpasim_wizard.check_config <hydra args...>is a new command which can be ran on login node to quickly sanity-check if the run configuration is valid in terms of syntax and scene settings.