Applied Intuition · Technical Report

TerraZero: Procedural Driving Simulation for Zero-Demonstration Self-Play at Scale

A fast object-level traffic simulator, a procedural scenario generator, and a compute-efficient self-play recipe that trains driving policies from scratch — with zero human demonstrations — and transfers them across cities and datasets.

Zhouchonghao Wu^1*, Akshay Rangesh^1*, Weixin Li¹, Wei-Jer Chang^1,2, Zachary Lee¹, Saeed Bonab¹, Tim Wang¹, Wei Zhan^1,2†

¹ Applied Intuition · ² UC Berkeley · ^* equal contribution · ^† corresponding author

Technical Report arXiv

@techreport{terrazero2026,
  title       = {TerraZero: Procedural Driving Simulation for Zero-Demonstration Self-Play at Scale},
  author      = {Wu, Zhouchonghao and Rangesh, Akshay and Li, Weixin and Chang, Wei-Jer and Lee, Zachary and Bonab, Saeed and Wang, Tim and Zhan, Wei},
  institution = {Applied Intuition},
  year        = {2026},
  type        = {Technical Report}
}

Contributions

A Fast, Feature-Rich Driving Simulator

A configurable C engine built for reinforcement learning at scale — high throughput without giving up realism. It models vehicles, pedestrians, cyclists, and high-fidelity trucks together, with real traffic rules and maps in 3D.

Hard Scenarios, Generated on Demand

A procedural generator that treats logged data as a starting point, not a ceiling. From real maps it composes an effectively unbounded supply of training scenarios — randomized starts and goals, reactive road users, and varied traffic-signal timing.

Driving Policies Trained From Scratch

The Zero in TerraZero names the training stance: zero human demonstrations. A compute-efficient self-play recipe learns by reinforcement signal alone — no imitation, no logged trajectories — and every result we report is trained this way. The resulting policies transfer zero-shot across cities and datasets, picking up region-specific habits such as left-hand traffic on their own.

Interactive Demo

Few Maps, Infinite Scenarios

A handful of real maps becomes an effectively unbounded stream of long-tail scenarios. Compose one with the controls and watch TerraZero drive it: choose the agent type (car, truck, or heterogeneous†), set a traffic condition, drop in road users and hazards, optionally dial in a robustness setting, and pick a viewpoint (agent or top-down) — the simulator procedurally generates the rest, and the selectors narrow to whatever it has on hand.

† Heterogeneous agents controlled via a unified policy will be released in work in preparation, to appear on arXiv.

▷

Scenario gallery

Enable JavaScript to browse rollouts across regions, traffic controls, road users, and camera views.

Simulator

Built for Speed

Large-scale reinforcement learning needs billions of environment steps. TerraZero is engineered for exactly that: a C simulation engine paired with GPU policy inference, delivering the throughput RL requires without trading away realism.

560K

steps/s · consumer GPU

1.3M

steps/s · server GPU

2.8M

steps/s · 8× server node

Agent steps per second, measured against object-level simulators that report comparable numbers. Hover a bar for the exact value.

Figure 1. Throughput across hardware tiers, in agent steps per second. Values marked with a dagger are reported by the original authors.

High-Fidelity Simulation

Speed never comes at the cost of realism. TerraZero models dynamics for pedestrians, cyclists, vehicles, and high-fidelity trucks; enforces real traffic rules including traffic lights and stop signs; and represents maps in 3D — so policies learn genuine compliance rather than workarounds for scripted flags.

Scenario Generation

The Situations That Matter Most

Real driving logs are overwhelmingly routine. The rare, high-stakes moments — dense merges, sudden cut-ins, near-miss crossings — are exactly what a driving policy most needs to practice, and exactly what logs contain least. TerraZero generates them on demand: from real maps it builds an effectively unlimited stream of challenging scenarios, with reactive traffic, hazards, and signal timing that change every episode.

Figure 2. Procedurally generated road users and hazards: construction and static obstacles, crossing and jaywalking pedestrians, and parked and collided vehicles.

Recipe & Results

State-of-the-Art Driving From Self-Play Alone

#1 94.19 nuPlan val14standard closed-loop

#1 70.87 InterPlaninteractive long tail

TerraZero trains driving policies entirely through self-play reinforcement learning — no human demonstrations, no logged trajectories. Because the simulator is so fast, the recipe trades sample efficiency for compute efficiency: it learns the hard cases from scratch, stays stable at scale, and runs a lightweight policy across many GPUs. One checkpoint, trained with one configuration, tops both the standard val14 benchmark and the interactive long-tail InterPlan suite — a pairing no prior planner achieves — and it transfers to new cities and datasets out of the box.

nuPlan val14

Top of the Standard Benchmark

Trained exclusively through self-play RL with zero human demonstrations, TerraZero posts the highest val14 score of any planner we compare against — ahead of the hand-engineered rule-based planners that have long topped this benchmark, the hybrids that graft a rule-based scorer onto a learned model, the imitation planners, and the rest of the RL family. The safety terms carry it: the best no-at-fault collision rate (99.02) and the best time-to-collision (95.62) in the field. Bars compare the strongest entry in each planner category; the radar breaks the composite into its eight components.

Planner	Type	Score	Ego Prog.	No AF-Coll.	Comfort	TTC	Drv. Dir.	Speed Lim.	Drv. Area	Making Prog.
TerraZero	RL	94.19	92.91	99.02	97.94	95.62	99.51	99.27	99.37	99.02
Gigaflow	RL	93.8	93.6	98.4	96.4	93.8	99.6	99.9	99.7	99.0
FlowDrive w/ guidance + PDM scoring	Hybrid	92.96	—	—	—	—	—	—	—	—
G2DP† + refinement	Hybrid	92.92	—	—	—	—	—	—	—	—
SPDM (N_p=15)	Rule	92.28	—	—	—	—	—	—	—	—
PDM-Closed	Rule	92.13	90.26	97.90	94.72	93.83	99.96	99.83	99.46	99.11
SPDM (N_p=60)	Rule	91.60	—	—	—	—	—	—	—	—
CaRL	RL	90.60	91.30	97.05	88.55	92.31	99.15	99.36	99.91	99.11
PLUTO	Hybrid	89.66	86.11	97.09	91.86	94.10	99.87	98.86	99.28	98.57
FlowDrive	IL	85.37	—	—	—	—	—	—	—	—
PlannerRFT	RL	84.46	—	—	—	—	—	—	—	—
G2DP	IL	83.91	—	—	—	—	—	—	—	—
Flow Planner	IL	83.31	—	—	—	—	—	—	—	—
Diffusion Planner*	IL	82.73	85.88	93.07	89.45	88.19	99.82	98.29	97.85	96.42
Log Replay (iLQR)	Replay	82.05	99.32	85.69	98.93	80.77	99.28	96.46	99.55	100.00
IDM	Rule	77.33	85.20	89.22	93.11	81.57	99.24	97.20	93.02	96.33
PlanTF	IL	76.14	77.21	95.21	93.56	90.61	99.33	98.50	96.33	89.53
Urban Driver	IL	53.05	87.17	72.45	99.28	66.37	96.87	81.66	83.36	94.28

Figure 3. Driving-policy evaluation on nuPlan val14 (closed-loop reactive), higher is better, showing the strongest entries from all four planner categories — rule-based, hybrid, imitation, and reinforcement learning. Radar axes run 0–100. An asterisk marks author-retrained variants, † the history-free G2DP variant with rule-based post-hoc refinement, and N_p the SPDM proposal budget. TerraZero is, to our knowledge, the first policy trained purely by large-scale self-play to outscore the dedicated nuPlan planners that top this benchmark.

InterPlan

State-of-the-Art on the Long Tail

The same checkpoint, with no retraining, also leads InterPlan's long-tail closed-loop suite — construction zones, accident sites, jaywalkers, nudging, overtaking, and dense lane changes — ahead of every rule-based, imitation, hybrid, and RL baseline, and ahead of the LLM-based planners, using no rule-based planner and no language model at inference. Prior planners trade one benchmark for the other: SPDM reaches 63.66 here only at its widest 60-proposal budget, which gives back val14, while the 15-proposal setting that peaks on val14 scores 42.00 on the long tail.

Planner	Type	Score	Ego Prog.	No AF-Coll.	Comfort	TTC	Drv. Dir.	Speed Lim.	Drv. Area	Making Prog.
TerraZero	RL	94.19	92.91	99.02	97.94	95.62	99.51	99.27	99.37	99.02
Gigaflow	RL	93.8	93.6	98.4	96.4	93.8	99.6	99.9	99.7	99.0
FlowDrive w/ guidance + PDM scoring	Hybrid	92.96	—	—	—	—	—	—	—	—
G2DP† + refinement	Hybrid	92.92	—	—	—	—	—	—	—	—
SPDM (N_p=15)	Rule	92.28	—	—	—	—	—	—	—	—
PDM-Closed	Rule	92.13	90.26	97.90	94.72	93.83	99.96	99.83	99.46	99.11
SPDM (N_p=60)	Rule	91.60	—	—	—	—	—	—	—	—
CaRL	RL	90.60	91.30	97.05	88.55	92.31	99.15	99.36	99.91	99.11
PLUTO	Hybrid	89.66	86.11	97.09	91.86	94.10	99.87	98.86	99.28	98.57
FlowDrive	IL	85.37	—	—	—	—	—	—	—	—
PlannerRFT	RL	84.46	—	—	—	—	—	—	—	—
G2DP	IL	83.91	—	—	—	—	—	—	—	—
Flow Planner	IL	83.31	—	—	—	—	—	—	—	—
Diffusion Planner*	IL	82.73	85.88	93.07	89.45	88.19	99.82	98.29	97.85	96.42
Log Replay (iLQR)	Replay	82.05	99.32	85.69	98.93	80.77	99.28	96.46	99.55	100.00
IDM	Rule	77.33	85.20	89.22	93.11	81.57	99.24	97.20	93.02	96.33
PlanTF	IL	76.14	77.21	95.21	93.56	90.61	99.33	98.50	96.33	89.53
Urban Driver	IL	53.05	87.17	72.45	99.28	66.37	96.87	81.66	83.36	94.28

Figure 4. InterPlan long-tail closed-loop reactive score (higher is better) on the official 80-scenario split, comparing the top contenders across planner categories. TerraZero uses no rule-based planner or language model at inference. On the full 335-scenario set the same checkpoint scores 71.31.

WOSAC

Sim-Agent Evaluation

As a sim agent, TerraZero is the strongest method that uses no demonstrations at all — no imitation and no reference policy — matching SPACeR and beating HR-PPO, which both derive from a reference policy trained on logged data. The demonstration-based CAT-K and the logged expert (dashed) are shown for reference.

WOSAC 2023

Planner	Type	Score	Ego Prog.	No AF-Coll.	Comfort	TTC	Drv. Dir.	Speed Lim.	Drv. Area	Making Prog.
TerraZero	RL	94.19	92.91	99.02	97.94	95.62	99.51	99.27	99.37	99.02
Gigaflow	RL	93.8	93.6	98.4	96.4	93.8	99.6	99.9	99.7	99.0
FlowDrive w/ guidance + PDM scoring	Hybrid	92.96	—	—	—	—	—	—	—	—
G2DP† + refinement	Hybrid	92.92	—	—	—	—	—	—	—	—
SPDM (N_p=15)	Rule	92.28	—	—	—	—	—	—	—	—
PDM-Closed	Rule	92.13	90.26	97.90	94.72	93.83	99.96	99.83	99.46	99.11
SPDM (N_p=60)	Rule	91.60	—	—	—	—	—	—	—	—
CaRL	RL	90.60	91.30	97.05	88.55	92.31	99.15	99.36	99.91	99.11
PLUTO	Hybrid	89.66	86.11	97.09	91.86	94.10	99.87	98.86	99.28	98.57
FlowDrive	IL	85.37	—	—	—	—	—	—	—	—
PlannerRFT	RL	84.46	—	—	—	—	—	—	—	—
G2DP	IL	83.91	—	—	—	—	—	—	—	—
Flow Planner	IL	83.31	—	—	—	—	—	—	—	—
Diffusion Planner*	IL	82.73	85.88	93.07	89.45	88.19	99.82	98.29	97.85	96.42
Log Replay (iLQR)	Replay	82.05	99.32	85.69	98.93	80.77	99.28	96.46	99.55	100.00
IDM	Rule	77.33	85.20	89.22	93.11	81.57	99.24	97.20	93.02	96.33
PlanTF	IL	76.14	77.21	95.21	93.56	90.61	99.33	98.50	96.33	89.53
Urban Driver	IL	53.05	87.17	72.45	99.28	66.37	96.87	81.66	83.36	94.28

WOSAC 2024

Planner	Type	Score	Ego Prog.	No AF-Coll.	Comfort	TTC	Drv. Dir.	Speed Lim.	Drv. Area	Making Prog.
TerraZero	RL	94.19	92.91	99.02	97.94	95.62	99.51	99.27	99.37	99.02
Gigaflow	RL	93.8	93.6	98.4	96.4	93.8	99.6	99.9	99.7	99.0
FlowDrive w/ guidance + PDM scoring	Hybrid	92.96	—	—	—	—	—	—	—	—
G2DP† + refinement	Hybrid	92.92	—	—	—	—	—	—	—	—
SPDM (N_p=15)	Rule	92.28	—	—	—	—	—	—	—	—
PDM-Closed	Rule	92.13	90.26	97.90	94.72	93.83	99.96	99.83	99.46	99.11
SPDM (N_p=60)	Rule	91.60	—	—	—	—	—	—	—	—
CaRL	RL	90.60	91.30	97.05	88.55	92.31	99.15	99.36	99.91	99.11
PLUTO	Hybrid	89.66	86.11	97.09	91.86	94.10	99.87	98.86	99.28	98.57
FlowDrive	IL	85.37	—	—	—	—	—	—	—	—
PlannerRFT	RL	84.46	—	—	—	—	—	—	—	—
G2DP	IL	83.91	—	—	—	—	—	—	—	—
Flow Planner	IL	83.31	—	—	—	—	—	—	—	—
Diffusion Planner*	IL	82.73	85.88	93.07	89.45	88.19	99.82	98.29	97.85	96.42
Log Replay (iLQR)	Replay	82.05	99.32	85.69	98.93	80.77	99.28	96.46	99.55	100.00
IDM	Rule	77.33	85.20	89.22	93.11	81.57	99.24	97.20	93.02	96.33
PlanTF	IL	76.14	77.21	95.21	93.56	90.61	99.33	98.50	96.33	89.53
Urban Driver	IL	53.05	87.17	72.45	99.28	66.37	96.87	81.66	83.36	94.28

TerraZero RL RL (Demonstration-derived) Demonstration-based

Figure 5. Sim-agent realism on the Waymo Open Sim Agents Challenge, evaluated zero-shot (higher is better). Solid RL bars use no demonstrations at all; shaded RL bars derive from a reference policy trained on logged data (SPACeR, HR-PPO); dashed bars are demonstration-based references (the logged Expert Demonstration and, on 2024, CAT-K). Left: the 2023 edition on the full Waymo validation split, against Gigaflow. Right: the 2024 edition on vehicles (shared 880-scenario subset). TerraZero uses no demonstrations and no reference policy, matching the realism of the demonstration-derived methods.

Terra Series

Empowering Physical AI Research

Learning to Drive Without Expert Demonstrations

TerraTransfer uses TerraZero's self-play planner as a teacher and distills its competence into an end-to-end camera policy — no human demonstrations required. It aligns the vision policy's latent space to the planner's vector representation, achieving 0.490 HD-Score on HUGSim and outperforming imitation baselines by +0.130.

Project Page → arXiv