Vision-Language-Action (VLA) models have made substantial progress in robotic manipulation, but their generalization remains constrained by the diversity and coverage of action-labeled robot demonstrations. Human videos provide much broader behavioral and environmental experience, while learning from them faces a fundamental trade-off between scalable data and explicit action grounding. We introduce ThinkWorld, a VLA learning framework that grounds action reasoning in world dynamics and strengthens it through reinforcement learning with world imagination. ThinkWorld represents each action thought as a Chain-of-Dynamics (CoD), a sequence of discrete motion tokens that captures the intermediate dynamics connecting the current world to a future state and serves as a common action representation across embodiments. We first initialize CoD with action-labeled demonstrations, then predict the future induced by each CoD and reinforce those whose predictions better match the observed outcome. This enables videos without action annotations to provide direct supervision from observed state transitions, allowing human and robot experience to refine the same reasoning space. Experiments in simulation and on real robots demonstrate effective human-to-robot transfer and strong generalization to unseen tasks and environments.
Figures & tables
Figure 1: We introduce ThinkWorld , a vision-language-action framework that learns transferable Chain-of-Dynamics (CoD) reasoning from diverse human and robot experiences. CoD captures action-induced dynamics across embodiments and is reinforced through future imagination, enabling robust generalization in the real world and skill composition guided by human videos.
Figure 2: ThinkWorld architecture and training stages. In stage (a), supervised fine-tuning (SFT) tokenizes robot actions and annotated human hand motions into a unified discrete motion space to ground Chain-of-Dynamics (CoD) reasoning, while CoD-conditioned imagination predicts the future visual feature. In stage (b), reinforcement learning (RL) uses the alignment between imagined and observed futures as a scalable reward for CoD reasoning over human and robot videos, strengthening action reasoning without motion annotations and improving guidance for continuous control.
Figure 3: Real-world out-of-distribution (OOD) generalization. Success rates for color matching, cup hanging, and sorting under in-domain and OOD conditions. OOD settings introduce new colors, objects, environments, or sorting trajectories, as illustrated below the plots. ThinkWorld achieves the highest OOD success rates across all three tasks, despite substantial drops for the baselines.
Figure 4: Composing atomic robot skills from a human demonstration. The top shows atomic robot skills, while the middle shows a human demonstration that provides the complete task composition, e.g., open → put-left → close. Guided by this human demonstration, ThinkWorld composes the corresponding robot skills and executes the complete task in a single rollout at the bottom .
LIBERO
LIBERO-PRO
Model
Params.
Spatial
Object
Goal
Long
Avg.
Object
Position
Semantic
Task
Environment
Avg.
OpenVLA ( Kim et al., 2024 )
7.5
84.7
88.4
79.2
53.7
76.5
93.0
0.0
97.3
0.0
68.0
51.7
π0 ( Black et al., 2024a )
3.2
96.8
98.8
95.8
85.2
94.2
90.5
0.0
90.5
0.0
38.8
44.0
π0.5 ( Intelligence et al., 2025 )
3.4
98.8
98.2
98.0
92.4
96.9
96.0
20.8
95.8
0.8
52.8
53.2
MolmoAct ( Lee et al., 2025 )
8.1
87.0
95.4
87.6
77.2
86.6
76.0
1.5
85.8
1.5
–
41.2
VLA-JEPA ( Sun et al., 2026a )
2.8
96.2
99.6
97.2
95.8
97.2
80.4
3.5
97.2
20.8
41.0
48.5 †
Table 1: Success rates (%) on LIBERO and LIBERO-PRO. LIBERO-PRO axis values are arithmetic means over Goal, Spatial, LIBERO-10, and Object. LIBERO-PRO baseline results are taken from the official repository unless marked with † , which denotes our reproduced result. Best and second-best results within each benchmark column are shown in bold and underlined , respectively.
Figure 6
Figure 7: Effect of Chain-of-Dynamics (CoD) on future prediction. We compare future DINO predictions under generated, shuffled, empty, and no-CoD conditions against the current and future ground-truth features. The generated CoD captures the future-state structure, whereas the other variants largely remain similar to the current state and fail to reflect the expected future change.
Figure 8: Attention analysis of CoD. We visualize attention from CoD tokens to the current observation. Across human and robot examples, CoD attends to action-relevant regions, including human hands, robot arms, and manipulated objects, highlighting its focus on motion-relevant cues.
Table 9
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Source
Episodes
Frames
Hours
Hz
Camera view
Action representation
YAM
32.2K
76.0M
704.1
30
Top + two wrists
30×14 D absolute joint
DROID
74.6K
17.8M
328.9
15
Exterior + wrist
15×8 D absolute joint
RoboCOIN
112.0K
82.3M
755.2
30/50
One or three views
30/50×(12/14) D absolute EEF
Bridge
53.2K
1.9M
105.2
5
Main
5×7 D delta EEF
RT-1
87.2K
3.8M
350.6
3
Main
3×7 D delta EEF
BC-Z
39.4K
5.5M
152.0
10
Main
10×7 D delta EEF
Appendix
Table 4: Training datasets and native observation/action contracts. K and M denote thousands and millions, respectively.
Setting
VLM SFT
Action SFT
RL
Initialization
Qwen3-VL-2B
VLM SFT
Action SFT
Trainable modules
Language model + future projector
Action expert
Language model
Objective
CoD/VQA CE +3Lfuture
Masked flow matching
GRPO/GDPO + weak SFT replay
Data
Robot + human motion; auxiliary VQA
Robot trajectories
Robot + human motion
Iterations
150k
50k (5k current-view + 45k history)
2k
Learning rate
VLM 10−5 ; projector 10−4
Expert 10−4→5×10−5
10−6
Appendix
Table 5: Training configuration . Batch is the effective global batch, and allocations report representative resources used for the runs.
Figure 9: BimanualYAM setup.
Color Matching
Cup Hanging
Sorting
Overall
Model
ID
OOD
ID
OOD
ID
OOD
Macro SR
Micro SR
Diffusion Policy
18/30 (60.0)
11/30 (36.7)
9/15 (60.0)
1/15 (6.7)
4/15 (26.7)
1/15 (6.7)
32.8
36.7
MolmoAct2
24/30 (80.0)
8/30 (26.7)
11/15 (73.3)
14/15 (93.3)
12/15 (80.0)
3/15 (20.0)
62.2
60.0
LaWAM
16/30 (53.3)
8/30 (26.7)
9/15 (60.0)
12/15 (80.0)
4/15 (26.7)
4/15 (26.7)
45.6
44.2
ThinkWorld (ours)
30/30 (100.0)
24/30 (80.0)
15/15 (100.0)
15/15 (100.0)
15/15 (100.0)
13/15 (86.7)
94.4
93.3
Appendix
Table 6: Full real-robot success-rate (SR) comparison. Each task is evaluated under in-distribution (ID) and out-of-distribution (OOD) settings; cells report successes/trials with SR (%) in parentheses. Task 1 comprises six ID and six OOD conditions ( 5 trials each); Tasks 2 and 3 comprise three of each. Macro SR is the unweighted mean of the six per-setting rates; Micro SR pools all 120 trials.
Group
Task
SR (%)
Google-VM
49.31
Google-VA
Pick Coke can
56.00
Drawer
33.33
Move-near
55.83
Put-in-drawer
2.12
Micro avg.
46.54
Appendix
Table 7: Out-of-the-box performance on SimplerEnv.
Method
Camera
Robot
Language
Light
Background
Noise
Layout
Avg.
OpenVLA-OFT ( Kim et al., 2024 )
56.4
31.9
79.5
88.7
93.3
75.8
74.2
69.6
π0 -FAST ( Black et al., 2024a )
65.1
21.6
61.0
73.2
73.2
74.4
68.8
61.6
DreamVLA ( Zhang et al., 2026 )
65.0
40.8
63.5
85.7
82.6
84.9
74.0
69.9
ThinkWorld (Ours)
46.9
70.5
73.5
90.7
90.0
62.7
79.6
71.7
Appendix
Table 8: Success rates (%) on LIBERO-Plus across seven perturbation dimensions. Only methods with published LIBERO-Plus results are listed. Higher is better.
Policy
p50
p95
Wall
Mem.
(ms)
(ms)
(s)
(GiB)
MolmoAct2
335.07
467.96
18.44
12.11
MolmoAct2-Think
879.37
1014.47
46.76
12.12
ThinkWorld
324.84
333.09
16.31
7.07
Appendix
Table 9: Inference efficiency. Per-call latency percentiles, total wall time, and peak GPU memory, measured at steady state over 50 consecutive eager-mode calls on a single NVIDIA H100 GPU. Lower is better for all metrics.
Figure 10: Real-robot experiment task rollout results. (a) Sample teleoperation demonstrations used for training. (b) In-domain rollouts. (c) Out-of-domain rollouts. Each panel shows qualitative rollout quality for every method, with frames sampled uniformly over an episode and time advancing left to right; the corresponding success rates are reported in Tab. 6 .
Figure 11: Real-robot experiment task rollout results. (a) Sample teleoperation demonstrations used for training. (b) In-domain rollouts. (c) Out-of-domain rollouts. Each panel shows qualitative rollout quality for every method, with frames sampled uniformly over an episode and time advancing left to right; the corresponding success rates are reported in Tab. 6 .
Figure 12: Real-robot experiment task rollout results. (a) Sample teleoperation demonstrations used for training. (b) In-domain rollouts. (c) Out-of-domain rollouts. Each panel shows qualitative rollout quality for every method, with frames sampled uniformly over an episode and time advancing left to right; the corresponding success rates are reported in Tab. 6 .
Figure 13: Real-Robot skill composition sample trajectory and rollout results. (a) Sample atomic robot trajectories for opening the oven (OPEN), placing the left or right object inside (PUT-L, PUT-R), and closing the oven (CLOSE). (b) Human demonstrations that combine these skills into open → close and open → put-L → close sequences. (c) Successful robot rollouts for the corresponding natural-language instructions. Frames in each row progress from left to right.
Figure 14: Real-Robot skill composition sample trajectory and rollout results. (a) Sample atomic robot trajectories for opening the upper or lower drawer (OPEN), placing the left or right object inside (PUT-L, PUT-R), and closing the drawer (CLOSE). (b) Human demonstrations that combine these skills into open → put-L → close and open → put-R → close sequences for the upper drawer, and an open → put-L → close sequence for the lower drawer. (c) Successful robot rollouts for the corresponding natural-language instructions. Frames in each row progress from left to right.
Figure 15: Qualitative future-prediction results. Each row is one episode. The first two columns show the current and ground-truth future RGB observations, and the next two show PCA projections of the ground-truth DINO features for those frames. The remaining five columns show the predicted future DINO features under different CoD conditions: with CoD, shuffled CoD, empty CoD, a CoD taken from a different (donor) episode, and without CoD. All PCA panels share one color mapping.
Figure 16: Qualitative future-prediction results. Each row is one episode. The first two columns show the current and ground-truth future RGB observations, and the next two show PCA projections of the ground-truth DINO features for those frames. The remaining five columns show the predicted future DINO features under different CoD conditions: with CoD, shuffled CoD, empty CoD, a CoD taken from a different (donor) episode, and without CoD. All PCA panels share one color mapping.
Figure 17: Additional attention visualizations. Attention maps from CoD and future-prediction tokens to the input observation. CoD attention concentrates on dynamic, task-relevant regions, while future-prediction attention is more broadly distributed across the scene.
Figure 18: One-second-ahead future predictions during closed-loop SimplerEnv rollouts. For each task, the upper row shows observations actually reached by the policy, while the lower row shows PCA-projected future representations predicted at the preceding replanning horizon. Each column pairs a prediction with the observation at its target time, so their scene structure should correspond. The close qualitative alignment in layout and task progress shows that the predictions track the states encountered during closed-loop execution. The initial frame has no preceding prediction.
Most Vision-Language-Action (VLA) models map observations directly to actions without explicit reasoning, limiting their capacity for reasoning-intensive long-horizon tasks. To address this, existing approaches adopt Chain-of-Thought (CoT) reasoning to enable subgoal decomposition and spatial anticipation. However, those methods lack a unified architecture for effective cross-modal reasoning and fail to explicitly include inverse reasoning ability based on the target state. We argue that manipulation planning naturally decomposes into prediction, anticipating the next visual state, and inverse dynamics, inferring the actions to reach it. Bridging both requires a unified autoregressive architecture that interleaves textual and visual reasoning in a single generation process. We propose ThinkingVLA, a generative model that realizes this decomposition within a unified Mixture-of-Transformers architecture. ThinkingVLA consists of a forward CoT that identifies the immediate subgoal and guides the visual forecasting; the predicted image then serves as the target state, grounding an inverse CoT that reasons about spatial relationships and action intent based on the predicted image; and the final action is generated conditioned on this full reasoning context. Extensive experiments on simulation and real-world benchmarks demonstrate that ThinkingVLA consistently outperforms state-of-the-art baselines, with particularly large gains on long-horizon manipulation tasks.
Tianyi Lu, Hui Zhang, Zijie Diao +8
Fudan University · Shanghai Innovation Institue · China Unicom +1
Vision-Language-Action (VLA) models have shown promising capabilities for embodied intelligence, but most existing approaches rely on text-based chain-of-thought reasoning where visual inputs are treated as static context. This limits the ability of the model to actively revisit the environment and resolve ambiguities during long-horizon tasks. We propose VLA-Thinker, a thinking-with-image reasoning framework that models perception as a dynamically invocable reasoning action. To train such a system, we introduce a two-stage training pipeline consisting of (1) an SFT cold-start phase with curated visual Chain-of-Thought data to activate structured reasoning and tool-use behaviors, and (2) GRPO-based reinforcement learning to align complete reasoning-action trajectories with task-level success. Extensive experiments on LIBERO and RoboTwin 2.0 benchmarks demonstrate that VLA-Thinker significantly improves manipulation performance, achieving 97.5% success rate on LIBERO and strong gains across long-horizon robotic tasks. Project and Codes: https://cywang735.github.io/VLA-Thinker/ .
Chaoyang Wang, Wenrui Bao, Sicheng Gao +5
University of Central Florida · University of Würzburg · University of Southern California +1
Building generalizable agents for diverse applications remains a fundamental challenge. While imitation learning-based policies succeed in specific training environments, they often fail to generalize to novel scenes and tasks. In this work, we propose World Action Planner, a robot planning system that leverages the reasoning capabilities of Vision-Language Models (VLMs) and the physical grounding of a multi-task pose-image conditioned world model. Our system enables an agent to propose initial action plans and iteratively refine them via optimization and search, reasoning over imagined world model rollouts. We demonstrate that our approach achieves superior performance across compositional tasks, new layouts, and zero-shot generalization scenarios, significantly outperforming state-of-the-art end-to-end policy models such as VLAs and WAMs. Project website at worldactionplanner.github.io