Registration closes October 18, 2026
New team registration closes on October 18, 2026. The submission deadline remains October 31, 2026.
01 Overview
Welcome to the AlpaSim End to End Closed Loop Challenge. This competition invites teams to build autonomous driving policies and compare them head-to-head in realistic closed-loop simulation, where each policy's decisions shape the future scene it must handle.
Autonomous driving research has made major progress, but it remains hard to compare policies across labs and companies in a realistic, reproducible way. Open-loop evaluation is useful, but it misses the compounding effects that make driving hard: a small planning error can change future observations, interactions, and risk.
AlpaSim provides a shared simulator, public development data, starter tools, baseline policies, and a common containerized submission interface. Organizer-managed evaluation workers run submissions on private held-out scenarios and publish leaderboard results with both a policy capability score and a safety metric.
Across both tracks, the goal is not just to crown a winner. We want to learn which policy families are robust under distribution shift, where they fail, and how the community can make AV evaluation more trustworthy.
02 Announcements & Upcoming Changes
Final-run selection and configuration
At least the top three teams in each track will be selected for final runs. Organizers may include additional teams depending on how close the leaderboard scores are.
The Harmonizer diffusion model will be disabled for initial runs. It will be enabled for PAI final runs and remain disabled for nuPlan final runs.
Each selected contestant will receive one additional evaluation run in PAI and two additional evaluation runs in nuPlan to reduce potential variance in the final results.
03 Competition Status & Known Issues
Current server is using
0bb4c4bfe10951ea5589aa8ea514e422cf3c3506
This is the current competition deployment on the
e2e_challenge branch.
Known issues
Public nuPlan navtest MAP route foldbacks: In 10 of
1,485 public scenarios, near-stationary position noise can make
MAP route generation select the wrong lane and fail its foldback
check before the driver runs. This does not affect the private
competition set. For local evaluation, use the recorded ego pose
yaw for MAP lane matching instead of heading inferred from
consecutive positions, or exclude the affected scenarios from
local comparisons. The fix is planned for public main;
e2e_challenge will remain unchanged during the competition.
Let us know on the AlpaSim GitHub issues page if you find anything.
Initial re-opening commit
cd713e0d0563352ce45b99d0a53a7173d581bde2
This was the initial re-opening deployment on the
e2e_challenge branch. Its configuration included the
settings below.
Official submission limit: three per team per UTC calendar month
Teams may make three official submissions per UTC calendar
month, shared across PAI and nuPlan. The allowance resets at
00:00 UTC on the first day of each month; unused submissions
do not carry forward. Accepted submissions count even if
evaluation fails. Use the CLI limits command to
check your team's remaining allowance.
Each warmup track has a separate allowance of three submissions per team in a rolling 30-day window. Warmup submissions do not consume the official allowance.
Longer PAI evaluation turnaround
PAI evaluations use the final, larger evaluation suite and may take nearly an order of magnitude longer to complete.
Tunable controller parameters
Teams can now provide values for tunable controller parameters at submission time.
Terms and Conditions Added
Per internal legal guidance, terms and conditions have been added for competitors. All registered teams may submit, but some teams may be ineligible for prizes. After registering and before submitting, please read and accept the terms through the challenge command-line tool.
Full camera coverage
PAI rollouts now include rear-facing cameras, giving drivers image coverage around the full ego vehicle rather than only its forward and side views.
Ego footprint in the driver API
The rollout specification now includes the ego vehicle bounding box and its pose relative to the ego rig, so drivers can account for the vehicle footprint used by evaluation.
Local leaderboard construction
New local evaluation utilities include a curated PAI validation split and reference runs, so teams can construct a local ranking and better understand the evaluation criteria. The local score scale is illustrative and is not directly comparable to the official leaderboard.
Local nuPlan closed-loop evaluation
Teams can now run the full public navtest suite
locally for the nuPlan track. It is a public development suite,
not the private official competition suite, and no nuPlan
reference bundle is included yet.
Known issues
nuPlan off-road scoring: Off-road calculations were not enabled for nuPlan evaluations. In addition, issue #180 identifies a problem with the off-road algorithm. Additional testing preceded the code update and redeployment. Organizers will rerun only the evaluation stage for existing nuPlan submissions.
Soft open used
6b24e0159e5f92a1d9150ddab81f933745e24758
This previous version is retained for reference. It was also
on the e2e_challenge branch.
Ground-truth violations in some scenes
Some soft-open scenes could produce violations even when following
the ground-truth path. This was fixed in the initial re-opening commit,
cd713e0d0563352ce45b99d0a53a7173d581bde2.
04 Registration
Registration is limited to teams affiliated with academic or industry organizations. To protect limited evaluation capacity and ensure fair participation, the team captain's verified Hugging Face primary email must be an organization address.
Registrations are reviewed to prevent duplicate or non-genuine accounts from circumventing team-level participation limits. Independent groups from the same organization are welcome: identify your lab, department, or group and briefly describe the project in the registration form.
Full challenge rules and submission limits are in the Contestant Guide.
05 Rules & Fair Play
The challenge is designed to compare independently developed policies under a shared, organizer-managed evaluation protocol. These rules protect that comparison and the limited evaluation capacity available to all teams.
Team participation
Each participant may belong to only one registered team. Submission limits apply at the team level across its authorized submitters; teams and participants must not coordinate submissions or share accounts to bypass those limits. Multiple independent teams from the same organization are welcome.
Permitted resources
Award-eligible nuPlan submissions may use only publicly available datasets, annotations, and model weights. PAI-AV submissions may use private resources when the team has the right to use them and provides a high-level disclosure of their use.
Evaluation integrity
Submissions must not use hidden scenario information, private route identifiers, or any information unavailable to the policy during evaluation. Attempts to exploit the leaderboard, simulator, or platform; attack the service; or otherwise undermine fair evaluation may result in disqualification.
Reproducibility and review
Submitted policies must follow the published container contract and evaluation constraints. Organizers may request code, logs, or method details for award candidates, disputed results, or suspected rule violations, and may invalidate non-reproducible submissions.
Technical reports and awards
Award eligibility requires a final valid submission and a concise technical report for the relevant track. Reports are limited to four pages, excluding references and optional appendices, and should use the AlpaSim technical report template, which uses an attributed adaptation of the NeurIPS 2026 layout. Innovation prizes will be determined through organizer review of these reports and may overlap with leaderboard placement awards.
Use the AlpaSim GitHub issues for questions, reproducible bugs, and rule clarifications. Material clarifications will be reflected here.
06 Challenge Tracks
The competition has two complementary tracks, covering both large-scale geographically diverse driving data and a lower-barrier entry point for teams working with established nuPlan-style workflows.
Track 1 - PAI-AV
Physical AI AV Track
The larger-scale setting for testing whether promising policies hold up as scenario diversity and long-tail coverage increase.
Competition scenes come from Denmark, France, Germany, Japan, South Korea, Spain, Sweden, the United Kingdom, and the United States.
Track 2 - nuPlan
nuPlan Track
A lower-barrier track for teams building on the widely used nuPlan ecosystem or NAVSIM-style development workflows.
The nuPlan track uses a private evaluation dataset.
07 Timeline
- 2026-06-15 Competition goes live Registration, public data, starter tools, and submission instructions open.
- 2026-09-09 Competition reopens Registration, submissions, and leaderboards are available again.
- 2026-09-15 Rules and submission format freeze Final rules, metric implementation, Docker base image, and public submission format freeze except for critical fixes.
- 2026-10-18 Registration cutoff New team registration closes.
- 2026-10-31 Public leaderboard closes Teams select one final valid container per track. Technical reports are due.
- 2026-11-15 (target) Target date for final results and winner announcements Final results and award decisions are targeted for release to participants on this date, subject to available compute and evaluation load.
- During NeurIPS 2026 Challenge results shared Winners and selected participants present results; organizers share final analysis and lessons learned.
08 Prizes
Each track will award two NVIDIA DGX Spark prizes: one for first place and one for an innovative solution.
The winners will receive a DGX Spark with an approximate retail value of USD $4,700. The prize must be accepted as awarded and is non-transferable, non-exchangeable, and has no cash value. NVIDIA is not responsible for any costs or taxes associated with prize acceptance. Teams marked (host) on the leaderboard are excluded from prizes.
DGX
PAI-AV Track
Two NVIDIA DGX Spark prizes awarded.
DGX
nuPlan Track
Two NVIDIA DGX Spark prizes awarded.
09 Documentation Links
Guide
Contestant Guide
Challenge setup, tracks, scoring, and participant workflow.
Kit
Starter Kit
Baseline files and examples for building an initial submission.
CLI
Submission CLI
Command-line tooling for packaging, validating, and submitting entries.
Requirements
Submission Image Requirements
Container image constraints and requirements for valid submissions.
Metrics
Capability Score Computation
Reference documentation for how policy capability scores are computed.
Framework
py123d_garage
Open-source training and closed-loop evaluation framework with pretrained baselines for the AlpaSim E2E Challenge.
Template
Technical Report Template
Four-page report skeleton for award-eligible final submissions, using an attributed NeurIPS 2026-style layout.
Support
Questions and Discussions
Use the public AlpaSim repository for bug reports, questions, and competition discussions.
10 Organizers
This competition is hosted by NVIDIA's Autonomous Vehicle Research Group, KE:SAI, and HKU.
Host
NVIDIA Autonomous Vehicle Research Group
Interdisciplinary NVIDIA Research team advancing vehicle autonomy across perception, prediction, planning, control, simulation, foundation models, and AI safety.
Host
KE:SAI
Non-profit open-science research lab advancing robust, safe, and reproducible physical AI, with a focus on world models, autonomy, and open self-driving technology.
Host
HKU
OpenDriveLab at The University of Hong Kong is a research group focused on embodied AI and autonomous driving. The lab develops open benchmarks, simulation infrastructure, world models, and end-to-end driving systems to support scalable, reproducible, and trustworthy physical AI research.