PHR-VLA: Planning Horizon Reasoning for Vision-Language-Action Models

Davood Soleymanzadeh; Kaidi Zhang; Zhiyuan Zhang; Bihao Zhang; Xiao Liang; Yu She; Minghui Zheng

cs.ROAug 27, 2026

PHR-VLA: Planning Horizon Reasoning for Vision-Language-Action Models

Authors: Davood Soleymanzadeh, Kaidi Zhang, Zhiyuan Zhang, Bihao Zhang, Xiao Liang, Yu She, Minghui Zheng

Organizations: J. Mike Walker ’66 Department of Mechanical Engineering, Texas A&M University, College Station, TX 77843, USA · Department of Industrial Engineering, Purdue University, West Lafayette, IN 47907, USA · Zachry Department of Civil and Environmental Engineering, Texas A&M University, College Station, TX 77843 USA

Abstract

Vision-language-action models (VLAs) have shown strong promise for general-purpose robotic manipulation by mapping language instructions and vision observations directly to actions. However, most VLAs primarily condition action prediction on current observations and lack an explicit mechanism for reasoning over future task dynamics, which is particularly important for fine-grained, contact-rich manipulation. We present PHR-VLA, a framework that enables planning-horizon reasoning in VLAs through privileged latent representations of future dynamics. PHR-VLA introduces a lightweight auxiliary future head that, during training, aligns the VLA's internal representations with latent dynamics extracted from future observations. Evaluation results demonstrate that local, contact-centric, patch-level latent dynamics supervision from the wrist camera improves success rate on LIBERO from 84.1% to 88.4% and on real-world disassembly tasks from 63.3% to 82.5%. Patch-level supervision from a third-person camera also improves performance on Meta-World from 56.70% to 57.8%. These results demonstrate that privileged latent dynamics alignment provides an effective training signal for improving anticipatory reasoning in VLA policies. Project website: https://davoodsz.github.io/PHR-VLA.github.io/

Figures & tables

Fig. 2: PHR-VLA Framework. Privileged Finetuning: In addition to the current observation $\mathbf{I}_{t}$ , language instruction $\mathbf{l}_{t}$ , and proprioceptive state $\mathbf{q}_{t}$ used by the standard VLA backbone, privileged planning horizon frames $\mathbf{I}_{t+1:t+H}$ are passed through a frozen vision encoder to compute latent-dynamics targets. The action-token latents produced by VLM backbone feed two heads: the standard flow matching action head, supervised by $\mathcal{L}_{\mathrm{VLA}}$ , and the auxiliary future head which is supervised by $\mathcal{L}_{\mathrm{Align}}$ against the latent dynamics targets. Efficient Inference: The action head runs exactly as in standard VLA backbone, so PHR-VLA adds zero latency or memory cost at deployment.

Method	Spatial $\uparrow$	Object $\uparrow$	Goal $\uparrow$	Long $\uparrow$	Average $\uparrow$
ACT [ 7 ]	30.7%	48.4%	02.0%	15.7%	24.2%
Diffusion Policy [ 5 ] ∗	78.3%	92.5%	68.3%	50.5%	72.4%
Octo [ 15 ] ∗	78.9%	85.7%	84.6%	51.1%	75.1%
DiT Policy [ 35 ] ∗	84.2%	96.3%	85.4%	63.8%	82.4%
OpenVLA [ 4 ] ∗	84.7%	88.4%	79.2%	53.7%	76.5%
SmolVLA [ 3 ]	86.0 %	97.7 %	86.7 %	66.0 %	84.1 %

TABLE I: Results on LIBERO. Success rates are reported across four task suites (300 trials per suite). Entries marked with ∗ are taken from [ 36 ] and may use different model sclaes, training setups, and computational budgets. The rest are trained under the same a single A100 GPU setting. Best results are bold , and second-best results are underlined .

Method	Easy $\uparrow$	Medium $\uparrow$	Hard $\uparrow$	VeryHard $\uparrow$	Average $\uparrow$
ACT [ 7 ]	68.2%	49.7%	28.9%	49.4%	49.0%
Diffusion Policy [ 5 ] ∗	23.1%	10.7%	01.9%	06.1%	10.5%
TinyVLA [ 6 ] ∗	77.6%	21.5%	11.4%	15.8%	31.6%
SmolVLA [ 3 ]	82.0 %	54.2 %	43.9 %	46.7 %	56.7 %
PHR-VLA (ours)	85.2 %	55.1 %	45.5 %	45.3 %	57.7 %

TABLE II: Results on Meta-World. Success rates are reported across all difficulty levels. Entries marked with ∗ are taken from [ 3 ] and may use different model sclaes, training setups, and computational budgets. The rest are trained under the same a single A100 GPU setting. Best results are bold , and second-best results are underlined .

Fig. 3: Ablation: Camera View. Success rate of SmolVLA and PHR-VLA with wrist-mounted, fixed third-person, and multi-camera auxiliary supervision.

Fig. 4: Ablation: Target Granularity. Success rate of PHR-VLA with various target supervision granularity across wrist-mounted, fixed-third person, and multi-camera.

Fig. 5: Ablation: Absolute Latent vs. Dynamics Latent Supervision. Success rate of PHR-VLA with planning-horizon absolute latents and dynamics latent supervision during training.

Fig. 6: Ablation:

\lambda

. Success rate of SmolVLA and PHR-VLA with various auxiliary loss weights (

\lambda

Method	View	Target	Type	Spatial $\uparrow$	Object $\uparrow$	Goal $\uparrow$	Long $\uparrow$	Average $\uparrow$
smolVLA [ 3 ]	–	–	–	86.0%	97.7%	86.7%	66.0%	84.1%
PHR-VLA (SigLIP)	Wrist	Patch	$\mathbf{y}$	88.7%	98.0%	92.4%	74.4%	88.4 %
	Wrist	Patch	$\mathbf{s}$	83.7%	98.7 %	91.7%	72.7%	86.7%
	Wrist	Mean	$\mathbf{y}$	87.0%	98.0%	91.0%	69.0%	86.2%
	Wrist	Mean	$\mathbf{s}$	84.0%	98.0%	89.0%	63.3%	83.6%
	Multicam	Patch	$\mathbf{y}$	87.3%	98.0%	90.7%	76.3 %	88.1 %

TABLE III: Ablation: Encoder Choice, SigLIP vs. JEPA on LIBERO. Success rates are reported across four task suites (300 trials per task) for SigLIP and JEPA encoders auxiliary supervision. “

\mathbf{y}

” denotes planning-horizon latent dynamics supervision, and “

\mathbf{s}

” denotes planning-horizon absolute latent supervision. Best results are bold , and second-best results are underlined .

Fig. 7: Disassembly Tasks. Contact-rich disassembly tasks for real-world evaluation of PHR-VLA and baseline policies.

Method	Target	Type	Easy $\uparrow$	Medium $\uparrow$	Hard $\uparrow$	Very Hard $\uparrow$	Average $\uparrow$
smolVLA [ 3 ]	–	–	82.0%	54.2%	43.9%	46.7%	56.7%
PHR-VLA (SigLIP)	Patch	$\mathbf{y}$	85.2%	55.1%	45.5%	45.3%	57.8%
	Patch	$\mathbf{s}$	84.2%	57.6 %	38.3%	47.3%	56.9%
	Mean	$\mathbf{y}$	84.3%	55.8%	50.0%	44.7%	58.7%
	Mean	$\mathbf{s}$	84.2%	57.6 %	38.3%	47.3%	56.9%
PHR-VLA (JEPA)	Patch	$\mathbf{y}$	82.0%	51.5%	48.3%	48.0 %	57.5%

TABLE IV: Ablation: Encoder Choice, SigLIP vs. JEPA on Meta-World. Success rates are reported across all difficulty levels for SigLIP and JEPA encoders auxiliary supervision. “

\mathbf{y}

” denotes planning-horizon latent dynamics supervision, and “

\mathbf{s}

” denotes planning-horizon absolute latent supervision. Best results are bold , and second-best results are underlined .

Task	Prompt	ACT [ 7 ]	Diffusion Policy [ 5 ]	SmolVLA [ 3 ]	PHR-VLA	PHR-VLA (JEPA)
Task I	Press the clip to release the object, remove it, then place it in the purple container.	19/30	21/30	26/30	28/30	29/30
Task II	Remove the object from the loose shaft, then place it in the purple container.	14/30	19/30	20/30	26/30	29/30
Task III	Slide the object inward, pull it out, then place it in the purple container.	10/30	14/30	24/30	22/30	26/30
Task IV	Twist the object 90 degrees and pull it out from the shaft, then place it in the purple container.	00/30	08/30	06/30	23/30	09/30
Total [#] $\uparrow$	–	43/120	62/120	76/120	99/120	93/120
Total [%] $\uparrow$	–	35.8%	51.7%	63.3%	82.5%	77.5%

TABLE V: Real-world Deployment. Success rate of PHR-VLA and baseline methods on real-world disassembly tasks. Best results are bold , and second-best results are underlined .

Fig. 8: PHR-VLA Deployment in Real-world. Examples of PHR-VLA deployment across the considered disassembly tasks, illustrating the policy execution across subsequent frames.

Explore similar work

ThinkingVLA: Interleaved Vision and Language Reasoning for Robotic Manipulation
Jun 16, 2026Tianyi Lu, Hui Zhang, Zijie Diao +8Long-Horizon Robotic Manipulation Language-Conditioned Robot Manipulation
Fudan University · Shanghai Innovation Institue · China Unicom +1
H-VLA: Hierarchical Vision-Language-Action Model with Key-Action Reasoning and Motion Planning in a Unified Action Space
Sep 19, 2026Xiongfeng Peng, Lu Xu, Yandong Wang +10Vision-Language-Action Models Language-Conditioned Robot Manipulation
Advanced Research Lab, Samsung R&D Institute China-Beijing (SRCB), China · Samsung AI Center, DS Division, South Korea
Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents
Jul 9, 2026Yixian Zhang, Huanming Zhang, Feng Gao +14Long-Horizon Robotic Manipulation Language-Conditioned Robot Manipulation
Tsinghua University · Striding AI · Purdue University +4

Jun 16, 2026, cs.ROJ/K move · Enter open · S save

ThinkingVLA: Interleaved Vision and Language Reasoning for Robotic Manipulation

Tianyi Lu, Hui Zhang, Zijie Diao, Junke Wang+7

Fudan University · Shanghai Innovation Institue · China Unicom+1

Long-Horizon Robotic Manipulation Language-Conditioned Robot Manipulation VLM Reasoning Vision-Language-Action Models Visually Grounded Reasoning Reasoning for VLA Models

Open paper AskPDF arXiv

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 \textbf{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.

Abstract

Explore similar work

ThinkingVLA: Interleaved Vision and Language Reasoning for Robotic Manipulation

H-VLA: Hierarchical Vision-Language-Action Model with Key-Action Reasoning and Motion Planning in a Unified Action Space

Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents