Vision-Language-Action (VLA) models have shown strong performance in robotic manipulation, but often struggle in long-horizon or out-of-distribution scenarios due to the lack of explicit mechanisms for multimodal reasoning and anticipating how the world will evolve under action. Recent works introduce textual chain-of-thought or visual subgoal prediction within VLA models to reason, but still fail to offer a unified human-like reasoning framework for joint textual reasoning, visual foresight, and action prediction. To this end, we propose HALO, a unified VLA model that enables embodied multimodal chain-of-thought (EM-CoT) reasoning through a sequential process of textual task reasoning, visual subgoal prediction for fine-grained guidance, and EM-CoT-augmented action prediction. We instantiate HALO with a Mixture-of-Transformers (MoT) architecture that decouples semantic reasoning, visual foresight, and action prediction into specialized experts while allowing seamless cross-expert collaboration. To enable HALO learning at scale, we introduce an automated pipeline to synthesize EM-CoT training data along with a carefully crafted training recipe. Extensive experiments demonstrate that: (1) HALO achieves superior performance in both simulated and real-world environments, surpassing baseline policy pi_0 by 34.1% on RoboTwin benchmark; (2) all proposed components of the training recipe and EM-CoT design help improve task success rate; and (3) HALO exhibits strong generalization capabilities under aggressive unseen environmental randomization with our proposed EM-CoT reasoning.
Figures & tables
Figure 1: Halo first performs textual reasoning and task planning, then predicts visual subgoals for fine-grained guidance, and finally generates actions conditioned on EM-CoT. This process is implemented using a Mixture-of-Transformers (MoT) architecture that integrates a multimodal understanding expert, a visual generation expert, and an action prediction expert together through shared self-attention.
Figure 2: Overview of EM-CoT Data Synthesis Pipeline. The pipeline converts raw robotic trajectories into EM-CoT data in three phases: (1) action primitives are extracted from robot proprioception via rule-based matching; (2) a VLM acts as an annotator to generate task plans, decompose trajectories into subtasks, and align each subtask with explicit textual reasoning; and (3) the terminal frame of each subtask is selected as a visual subgoal image, producing structured embodied multimodal chain-of-thought supervision.
Figure 3
Task
π0
RDT-1B
Diffusion Policy
Halo -w/o EM-CoT
Halo
Easy
Hard
Easy
Hard
Easy
Hard
Easy
Hard
Easy
Hard
Adjust Bottle
90
56
81
75
97
0
99
2
97
9
Beat Block Hammer
43
21
77
37
42
0
95
3
96
11
Blocks Ranking RGB
19
5
3
0
0
0
85
3
94
7
Blocks Ranking Size
7
1
0
0
1
0
37
0
58
2
Click Alarmclock
63
11
61
12
61
5
85
10
83
14
Table 1: Quantitative results in simulation on RoboTwin 2.0. Each task is evaluated 100 times under the Easy (Clean) and Hard (Domain-randomized) settings, and the average success rate across 50 tasks is reported. The complete results are provided in Appendix D .
Panel A: Training Recipe
Panel B: EM-CoT
Setting
Easy
Hard
Setting
Easy
Hard
Full
75.3
21.2
Halo
80.5
26.4
w/o V
58.2
10.5
w/o T
77.8
18.3
w/o V+T
42.9
3.9
w/o V
76.1
22.5
w/o V+T+A
32.4
0.0
w/o V & T
75.3
21.2
Table 2: Ablation studies on training recipe and EM-CoT. Panel A: training recipe ablation (V/T/A denote visual generation, textual VQA, and action prediction data. Note that in the Full setting, we pre-train the model with V&T&A datasets and then fine-tune it on the official RoboTwin 2.0 dataset). Panel B: EM-CoT ablation (V = visual subgoal images, T = textual reasoning).
Figure 6
Figure 7: Quantitative results of real-world experiments on both the basic setting and the generalization setting. Each task is evaluated over 50 trials, and the success rate is reported.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
Parameter
Value
architectures
[”Qwen2ForCausalLM”]
num_attention_heads
12
attention_dropout
0.0
num_hidden_layers
28
bos_token_id
151643
num_key_value_heads
2
eos_token_id
151645
rms_norm_eps
1×10−6
hidden_act
”silu”
rope_scaling
null
hidden_size
1536
rope_theta
1000000.0
Appendix
Table 3: Qwen2.5-1.5B-Instruct Model Configuration Parameters
Configuration
Pre-training
Fine-tuning
Base Architecture
Qwen2.5-1.5B × 3 Experts
Total Parameters
≈ 4.5B
Optimizer
AdamW
Learning Rate
1×10−4
5×10−5
Learning Rate Schedule
Constant
Constant
Weight Decay
0.0
0.0
Appendix
Table 4: Training Hyperparameters. Detailed configurations for the two-stage training process of Halo .
Task
π0
RDT-1B
Diffusion Policy
Halo -w/o CoT
Halo -w/ CoT
Easy
Hard
Easy
Hard
Easy
Hard
Easy
Hard
Easy
Hard
Adjust Bottle
90
56
81
75
97
0
99
2
97
9
Beat Block Hammer
43
21
77
37
42
0
95
3
96
11
Blocks Ranking RGB
19
5
3
0
0
0
85
3
94
7
Blocks Ranking Size
7
1
0
0
1
0
37
0
58
2
Click Alarmclock
63
11
61
12
61
5
85
10
83
14
Appendix
Table 5: Complete Result on RoboTwin
Figure 8: Attention map during pre-training phase.
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
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
Improving embodied reasoning in multimodal-large-language models (MLLMs) is essential for building vision-language-action models (VLAs) on top of them to readily translate multimodal understanding into low-level actions. Accordingly, recent work has explored enhancing embodied reasoning in MLLMs through supervision of vision-question-answering type. However, these approaches have been reported to result in unstable VLA performance, often yielding only marginal or even negative gains. In this paper, we propose a more systematic MLLM training framework RoboAlign that reliably improves VLA performance. Our key idea is to sample action tokens via zero-shot natural language reasoning and refines this reasoning using reinforcement learning (RL) to improve action accuracy. As a result, RoboAlign bridges the modality gap between language and low-level actions in MLLMs, and facilitate knowledge transfer from MLLM to VLA. To validate the effectiveness of RoboAlign, we train VLAs by adding a diffusion-based action head on top of an MLLM backbone and evaluate them on major robotics benchmarks. Remarkably, by performing RL-based alignment after SFT using less than 1% of the data, RoboAlign achieves performance improvements of 17.5%, 18.9%, and 106.6% over SFT baselines on LIBERO, CALVIN, and real-world environments, respectively.