RISE: Adaptive Imagination for World Action Models

Lu, Hongbo; Yao, Liang; He, Chenghao; Han, Hao; Liu, Fan; Liao, Wenlong; He, Tao; Peng, Pai

Technical Report · July 31, 2026

RISE: Adaptive Imagination for World Action Models

Hongbo Lu^1,2,*, Liang Yao^1,3,*, Chenghao He^1,*, Hao Han^1,*, Fan Liu³, Wenlong Liao^1,†, Tao He¹, Pai Peng^1,†,‡

¹COWARobot Co. Ltd ²Shanghai Jiao Tong University ³Hohai University
^*Equal Contribution · ^†Corresponding Author · ^‡Project Lead
Correspondence: volans.liao@cowarobot.com, pengpai@cowarobot.com

Code Data DAWN BibTeX

Comparison of fixed and adaptive imagination strategies — From fixed imagination strategies to RISE's scene-adaptive Roll/Stop scheduling.

Abstract

World Action Models improve planning by incorporating future world evolution into action generation, yet existing methods allocate a fixed imagination budget to every scene. RISE (Refining Imagination through SElective Rollout) makes sequential Roll/Stop decisions according to the expected planning benefit of continued rollout.

At each step, a Latent Evaluator estimates the risk revealed by the current prefix and how much planning could improve if imagination continues. A Rollout Gate weighs this expected benefit against additional computation cost. To overcome the single-future limitation of factual driving logs, CounterDrive supplies diverse counterfactual outcomes, risk levels, verified incident onsets, and causal annotations.

Experiments on NAVSIM and nuScenes show that RISE achieves the best overall planning performance while reducing unnecessary rollout. Transfer experiments further support its use as a plug-in scheduler across WAM architectures.

Adaptive Imagination

Future Planning Gain

Measures how much the planning score could improve by continuing from the current latent prefix.

Latent Evaluator

Predicts a Risk Profile for revealed hazards and a gain profile for the remaining rollout choices.

Rollout Gate

Balances predicted planning gain against computation cost and makes a fresh Roll/Stop decision at every step.

Overview of the RISE training and inference framework — RISE trains future prediction, variable-prefix planning, risk and gain estimation, and a cost-aware Rollout Gate.

CounterDrive

CounterDrive augments selected factual scenes with counterfactual 10-second driving videos. Human annotators verify ego-motion consistency, identify incident onset, mark generation distortions, and categorize causal behavior. Accepted clips supervise future prediction, while verified factual-counterfactual pairs provide temporally localized risk supervision.

5,013NAVSIM train clips

1,000NAVSIM test clips

2,432nuScenes train clips

511nuScenes test clips

Results

91.5NAVSIM v1 PDMS

Best overall planning performance, improving the strongest baseline by 0.8 points.

90.8NAVSIM v2 EPDMS

Ranks first or ties for first on seven of nine component metrics.

0.31 mnuScenes Avg. L2

State-of-the-art trajectory accuracy with a 0.10 average collision rate.

2.40Avg. Rollout

Adaptive scheduling reaches 90.8 EPDMS at 287.429 ms latency.

nuScenes Benchmark

Method	L2 (m) ↓				Collision Rate ↓
Method	1s	2s	3s	Avg.	1s	2s	3s	Avg.
BEV-Planner	0.30	0.52	0.83	0.55	0.10	0.37	1.30	0.59
LAW	0.26	0.57	1.01	0.61	0.14	0.21	0.54	0.30
World4Drive	0.23	0.47	0.81	0.50	0.02	0.12	0.33	0.16
WorldRFT	0.21	0.44	0.76	0.47	0.10	0.11	0.23	0.15
DAWN	0.17	0.31	0.52	0.33	0.00	0.10	0.23	0.11
RISE	0.16	0.29	0.49	0.31	0.00	0.11	0.20	0.10

NAVSIM v1 Benchmark

Method	NC ↑	DAC ↑	EP ↑	C ↑	TTC ↑	PDMS ↑
DrivingGPT	98.9	90.7	79.7	95.6	94.9	82.4
LAW	97.4	93.3	78.8	100	91.9	83.8
World4Drive	97.4	94.3	79.9	100	92.8	85.1
Epona	97.9	95.1	80.4	99.9	93.8	86.2
DriveVLA-W0	98.4	95.3	80.9	100	95.2	87.2
PWM	98.6	95.9	81.8	100	95.4	88.1
DreamerAD	98.0	97.2	83.1	100	94.3	88.7
DriveLaW	99.0	97.1	81.3	100	96.7	89.1
Drive-JEPA	98.7	96.2	82.9	100	95.5	89.0
DAWN	98.7	95.9	84.3	100	96.0	89.1
EponaV2	98.6	97.9	84.8	100	95.7	90.4
DriveFuture	98.8	99.1	95.4	100	84.2	90.7
RISE	99.1	97.7	98.3	100	98.6	91.5

NAVSIM v2 Benchmark

Method	NC ↑	DAC ↑	DDC ↑	TL ↑	EP ↑	TTC ↑	LK ↑	HC ↑	EC ↑	EPDMS ↑
DAWN	97.3	92.0	99.1	99.7	87.4	96.6	96.0	98.3	85.5	83.2
DreamerAD	98.0	97.2	99.5	99.8	87.8	97.4	97.5	98.3	72.4	85.1
DriveLaW	98.7	96.9	99.6	99.8	87.5	98.3	97.6	98.4	77.4	88.6
EponaV2	98.5	97.4	99.5	99.9	87.9	98.1	97.7	98.2	77.4	88.9
Latent-WAM	98.1	97.3	99.6	99.8	87.7	97.3	97.6	98.1	87.3	89.3
DriveFuture	98.8	99.1	99.6	99.9	86.6	98.4	96.4	98.3	74.8	89.9
RISE	99.1	97.7	99.7	99.9	87.8	98.7	98.0	98.4	87.4	90.8

Why Adaptive Rollout?

Distribution and performance across preferred rollout depths — Different scenes prefer different rollout depths; no fixed budget is uniformly optimal.

Representative scenes grouped by preferred rollout depth — Dense traffic and involved interactions generally benefit from deeper imagination.

Key Components

Scheduler	CounterDrive	EPDMS ↑	PDMS ↑
-	-	88.9	89.7
-	✓	89.8	90.5
✓	-	90.4	91.2
✓	✓	90.8	91.5

Adaptive Rollout Strategy

Method	Avg. Rollout	Latency (ms)	EPDMS ↑
Random Stop	2.03	264.075	89.5
Latent Margin	2.98	308.532	89.7
Scheduler	2.40	287.429	90.8

Qualitative Results

Qualitative trajectory comparison across adaptive rollout depths — RISE remains aligned with human trajectories and feasible lane geometry across scene-dependent rollout depths.

BibTeX

@misc{lu2026rise,
  title={RISE: Adaptive Imagination for World Action Models},
  author={Hongbo Lu and Liang Yao and Chenghao He and Hao Han and Fan Liu and Wenlong Liao and Tao He and Pai Peng},
  year={2026},
  url={https://cowarobot-ai.github.io/RISE/}
}