EVA: Aligning Video World Models with Executable Robot Actions via Inverse Dynamics Rewards

Ruixiang Wang; Qingming Liu; Yueci Deng; Guiliang Liu; Zhen Liu; Kui Jia

cs.ROMar 18, 2026

EVA: Aligning Video World Models with Executable Robot Actions via Inverse Dynamics Rewards

Authors: Ruixiang Wang, Qingming Liu, Yueci Deng, Guiliang Liu, Zhen Liu, Kui Jia

Organizations: The Chinese University of Hong Kong, Shenzhen · DexForce Technology Co., Ltd.

Abstract

Video generative models offer a promising basis for robotic planning by predicting future visual rollouts from observations and task instructions. However, their training objectives typically lack explicit constraints on robot executability. Generated rollouts can exhibit embodiment-related artifacts, including arm deformation and temporal discontinuities, while the actions decoded from these rollouts may be unstable or violate robot motion limits. We term this mismatch between generated robot behavior and embodiment-specific execution requirements the executability gap. We introduce Executable Video Alignment (EVA), a post-training framework that uses action-space feedback to improve both the visual integrity of generated robot motion and the feasibility of decoded actions. EVA trains an inverse dynamics model (IDM) on robot demonstrations and freezes it to decode actions from generated videos. We compute a reward from these actions that penalizes non-smooth motion and violations of velocity and acceleration limits, and use it to guide reinforcement-learning updates to the video generator. By feeding decoded-action constraints back into video generation, EVA links visual and action-level improvements through a shared alignment signal. Experiments on RoboTwin, LIBERO, and a real bimanual robot show that EVA reduces embodiment-related visual artifacts and improves task success when executing IDM-decoded actions. Further experiments show that video models fine-tuned with EVA generate more effective video--action pairs for training downstream VLAs.

Figures & tables

Figure 1: Overview of Executable Video Alignment (EVA). (a) EVA post-trains a video generator using rewards computed from actions decoded by a frozen inverse dynamics model (IDM), encouraging smoother, embodiment-consistent motion. (b) The aligned model supports video generation and video-based action planning, and can supply video–action pairs for downstream VLA training. The examples compare generated videos before and after alignment.

Figure 2: IDM-decoded actions from generated videos with and without visual artifacts. Plots show six left-arm joint-position commands and one gripper command, ordered from base to gripper. Top: coherent robot motion with smooth decoded commands. Bottom: visual artifacts (red boxes) accompanied by command fluctuations, especially in the distal joint and gripper (dimensions 6–7).

Figure 3: Qualitative video comparison on two tasks. Rows show EVA, Base, and Vidar; red circles highlight arm artifacts in Base and Vidar, while EVA shows more coherent robot geometry.

Method	Kinematic	Interaction	Instruction	Perfect
Vidar (Wan2.2)	67.6	66.7	87.6	62.9
Base	70.5	83.3	90.5	68.1
EVA	91.4	86.2	89.5	83.8

Table 1: Visual planning quality across 210 prompts, including Vidar ( Feng et al., 2025 ) . Entries are human-rated pass rates (%); Perfect requires all criteria and visually depicted task completion.

Figure 4: Real-robot deployment. Left: Generated video plans ; Right: corresponding robot executions.

Method	RoboTwin 2.0					LIBERO
Method	Place Mouse	Place Obj. Stand	Turn Switch	$\cdots$	Mean (%) 21 tasks	Spatial	Goal	Object
Base	10/50	28/50	23/50	$\cdots$	45.81	46.5	49.5	63.5
EVA	15/50	33/50	30/50	$\cdots$	52.29	53.5	57.0	70.0
$\Delta$	+5	+5	+7	$\cdots$	+6.48	+7.0	+7.5	+6.5

Table 2: Direct execution success. RoboTwin reports successes/50 trials for three representative tasks and mean success (%) over all 21 tasks under the clean setting. LIBERO reports success rates (%) over 200 trials per suite.

\Delta

: EVA minus Base, in counts for individual RoboTwin tasks and percentage points otherwise.

Figure 5: Real-robot success (%) on seen and OOD tasks with 20 trials each: ACT ( Zhao et al., 2023 ) ,

\pi_{0}

( Black et al., 2025 ) , Vidar ( Feng et al., 2025 ) , and GE-Act ( Liao et al., 2025 ) . Shaded groups show means; dashes denote unreported ACT OOD results.

Data source	Click Bell	Place A2B Right	Place Container Plate	Mean
Base-generated data	92	18	53	54.3
EVA-generated data	98	26	67	63.7

Table 3: Downstream

\pi_{0.5}

success after fine-tuning on Base- or EVA-generated data for three RoboTwin tasks. Entries are success rates (%) over 100 trials per task.

Method	$G_{\mathrm{test}}$	Success (%)
Auxiliary training (21 tasks)
Base	1	45.81
IDM-aux SFT	1	50.29
EVA	1	52.29
Inference-time controls (3 tasks)
Base	1	53.0

Table 4: Ablations and alternative strategies on RoboTwin. (a) Auxiliary training uses 21 tasks; inference-time controls use a separate three-task evaluation. Base + Smooth smooths decoded actions;

G_{\mathrm{test}}

denotes candidates per planning round. (b) Reward ablation uses another three-task evaluation. (c) EVA applied to Vidar ( Feng et al., 2025 ) on OpenLaptop. All rates are percentages.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1: Common failure modes observed in unaligned video world models during real-world execution. Implausible kinematics: violations of rigid-body consistency, including (a) morphological deformation, (b) ambiguous joint articulation, and (c) temporal discontinuity. Wrong contact: physically inconsistent object interaction. Incorrect goal: failure to make progress toward the instruction-conditioned objective.

Figure S2: Generated video comparison on LIBERO. Base is shown in the top row and EVA in the bottom row. The left four columns show placing the soup can in the basket (LIBERO-Object), and the right four show opening the middle drawer (LIBERO-Goal). Panel (a) samples frames 12, 24, 36, and 48 in chronological order from the third generated chunk of each model. Panel (b) shows representative Base artifacts from an earlier chunk and EVA interaction stages from subsequent planning rounds, rather than temporally matched frames. Red boxes highlight visible arm deformation in Base rollouts.

Figure S3: Relationship between reward scores and rollout quality. (a) Reward scores grouped by whether the generated rollout contains visible embodiment-related artifacts. (b) Reward scores grouped by downstream execution outcomes in simulation. Results are shown for three representative tasks, with 10 rollouts per group for each task. Each point represents one rollout, and box plots summarize the reward distributions.

Figure S4: Training reward curves in simulation and real-robot experiments. Rewards are normalized to $[0,10]$ for visualization. The blue curve shows the normalized reward, while the orange curve shows its exponential moving average (EMA). In both settings, the reward exhibits a generally increasing trend during alignment, showing the evolution of the optimization objective.

Figure S5: High-reward failure cases during GRPO post-training.

Inverse Dynamics Model	Test Accuracy	Test Success Rate
Spatial softmax	0.9864	89.52%
Global average pooling	0.7738	84.29%

Appendix

Table S1: Ablation study on the IDM architecture. Test accuracy requires all 14 action dimensions to have absolute errors below

0.05

rad for arm joints and

0.5

for

0/1

gripper commands simultaneously. Execution success is measured in simulation.

Video planner	ResNet IDM-A	ViT IDM-B
Base	45.8	43.5
EVA	52.3	49.7
$\Delta$	+6.5	+6.2

Appendix

Table S2: Cross-architecture IDM evaluation. Entries report execution success rates (%);

\Delta

denotes the improvement in percentage points. ResNet IDM-A is used for reward computation, while ViT IDM-B is used only for evaluation.

Task	ACT	DP	RDT	$\pi_{0}$	Base	EVA
Click Alarmclock	16/50	31/50	31/50	32/50	48/50	50/50
Click Bell	29/50	27/50	40/50	22/50	42/50	48/50
Handover Block	21/50	5/50	23/50	23/50	0/50	0/50
Handover Mic	43/50	27/50	45/50	49/50	1/50	8/50
Move Pillbottle Pad	0/50	1/50	4/50	11/50	8/50	16/50
Move Stapler Pad	0/50	1/50	1/50	0/50	0/50	1/50

Appendix

Table S3: RoboTwin clean-setting results on 21 tasks, comparing ACT ( Zhao et al., 2023 ) , DP ( Chi et al., 2023 ) , RDT ( Liu et al., 2024 ) , and

\pi_{0}

( Black et al., 2025 ) . Baseline entries are rounded /50 equivalents of published success rates; Base and EVA report observed successes out of 50 trials. Baseline averages are computed from the rounded entries.

Task Name	Arm Type	Task Description
StackBowl	Dual-Arm	The robot uses both arms to grasp two bowls from the two sides and stack them onto the target bowl placed at the center of the table.
HangCable	Single-Arm	The robot uses a single arm to grasp a black cable and hang it onto the designated rack.
Place2Basket	Single-Arm	The robot uses a single arm to pick up an object from the tabletop and place it into a nearby basket.
Place2Tray	Single-Arm	The robot uses a single arm to grasp an object from the table and place it onto the tray.
FoldTowel	Dual-Arm	The robot uses both arms to grasp the two bottom corners of a towel and fold it upward.
PlaceBlock (OOD)	Single-Arm	The robot uses a single arm to pick up a block and place it into the bowl with the specified color and position.

Appendix

Table S5: Real-world evaluation tasks used in our experiments.

Figure S7: EVA video generation on out-of-distribution tasks. Each row shows a synthesized video sequence of six evenly sampled frames, where the first frame serves as the input conditioning image and the instructions below correspond to the task prompts.

Figure S8: EVA rollouts conditioned on synthetic initial scenes. Each row shows a generated video sequence whose first frame is supplied by the text-to-image generator.

World Models for Robotics

World models for robotics predict how scenes and robot states change under actions to support planning, control, and learning.

See all 384 papers

RoboAlign-R1: Distilled Multimodal Reward Alignment for Robot Video World Models
May 5, 2026Hao Wu, Yuqi Li, Yuan Gao +10Multimodal Reward Modeling World Model Evaluation
Tencent · Tsinghua University · Department of Computer Science, City University of New York +3
DVG-WM: Disentangled Video Generation Enables Efficient Embodied World Model for Robotic Manipulation
Jun 30, 2026Ziyu Shan, Zhenyu Wu, Xiaofeng Wang +2Video Generation Robotic Manipulation
Nanyang Technological University, Singapore · Beijing University of Posts and Telecommunications, Beijing, China · GigaAI
GEM-4D: Geometry-Enhanced Video World Models for Robot Manipulation
May 20, 2026Kaichen Zhou, Yuzhen Chen, Fangneng Zhan +8Physical Consistency in Video Generation Robot Motion Generation
Harvard AI and Robotics Lab, Harvard University · Media Lab and EECS, MIT · MIT-IBM Watson AI Lab +1

May 5, 2026J/K move · Enter open · S save

RoboAlign-R1: Distilled Multimodal Reward Alignment for Robot Video World Models

Hao Wu, Yuqi Li, Yuan Gao, Fan Xu+9

Tencent · Tsinghua University · Department of Computer Science, City University of New York+3

Multimodal Reward Modeling World Model Evaluation Temporal Consistency in Video Generation Video World Models Robot Learning from Video

Open paper AskPDF arXiv

Existing robot video world models are typically trained with low-level objectives such as reconstruction and perceptual similarity, which are poorly aligned with the capabilities that matter most for robot decision making, including instruction following, manipulation success, and physical plausibility. They also suffer from error accumulation in long-horizon autoregressive prediction. We present RoboAlign-R1, a framework that combines reward-aligned post-training with stabilized long-horizon inference for robot video world models. We construct RobotWorldBench, a benchmark of 10,000 annotated video-instruction pairs collected from four robot data sources, and train a multimodal teacher judge, RoboAlign-Judge, to provide fine-grained six-dimensional evaluation of generated videos. We then distill the teacher into a lightweight student reward model for efficient reinforcement-learning-based post-training. To reduce long-horizon rollout drift, we further introduce Sliding Window Re-encoding (SWR), a training-free inference strategy that periodically refreshes the generation context. Under our in-domain evaluation protocol, RoboAlign-R1 improves the aggregate six-dimension score by 10.1% over the strongest baseline, including gains of 7.5% on Manipulation Accuracy and 4.6% on Instruction Following; these ranking improvements are further supported by an external VLM-based cross-check and a blinded human study. Meanwhile, SWR improves long-horizon prediction quality with only about 1% additional latency, yielding a 2.8% gain in SSIM and a 9.8% reduction in LPIPS. Together, these results show that reward-aligned post-training and stabilized long-horizon decoding improve task consistency, physical realism, and long-horizon prediction quality in robot video world models.

Abstract

World Models for Robotics

Explore similar work

RoboAlign-R1: Distilled Multimodal Reward Alignment for Robot Video World Models

DVG-WM: Disentangled Video Generation Enables Efficient Embodied World Model for Robotic Manipulation

GEM-4D: Geometry-Enhanced Video World Models for Robot Manipulation