Source-linked AI summary
OpenHelix: A Short Survey, Empirical Analysis, and Open-Source Dual-System VLA Model for Robotic Manipulation
Can Cui, Pengxiang Ding, Wenxuan Song, Shuanghao Bai, Xinyang Tong, Zirui Ge, Runze Suo, Wanqi Zhou, Yang Liu, Bofang Jia, Han Zhao, Siteng Huang, Donglin Wang
TL;DR
Dual-system VLAs face unresolved design and deployment challenges involving efficient real-time execution, model selection, information flow, and system integration. The paper surveys existing architectures, systematically examines core design choices, and presents a simple dual-system approach using prompt tuning and auxiliary tasks. Its reported conclusions favor integrating upper and lower model layers, adding auxiliary tasks, and using sparse asynchronous inference, while full open-source reproduction and real-world deployment remain limited.
Problem
Dual-system VLA design must balance the generalization of large multimodal models with the efficient action execution of lightweight policies, while real-time deployment remains difficult because of model size and resource demands.
Method
The paper surveys and compares dual-system architectures, then evaluates design choices involving model integration, prompt tuning, auxiliary tasks, and asynchronous inference.
Results
Integrating upper and lower model layers improves language generalization, auxiliary tasks enhance standard and generalization performance, and one asynchronous inference step leaves general-task performance largely unchanged.
Takeaways & Limitations
Prompt tuning with an auxiliary task provides a more robust latent embedding that integrates visual and textual information while avoiding direct MLLM fine-tuning.
Takeaways & Limitations
The project remains short of a full open-source reproduction of Helix, and direct VLA deployment is constrained by model size, resource demands, domain shift, and catastrophic forgetting.
Abstract
from arXiv · showhide
Dual-system VLA (Vision-Language-Action) architectures have become a hot topic in embodied intelligence research, but there is a lack of sufficient open-source work for further performance analysis and optimization. To address this problem, this paper will summarize and compare the structural designs of existing dual-system architectures, and conduct systematic empirical evaluations on the core design elements of existing dual-system architectures. Ultimately, it will provide a low-cost open-source model for further exploration. Of course, this project will continue to update with more experimental conclusions and open-source models with improved performance for everyone to choose from. Project page: https://openhelix-robot.github.io/.
1. Short Survey
Dual-system VLAs combine a slower, generalizing System 2-like model with a faster System 1-like policy, but their designs differ across modalities, latent representations, training, and integration. The survey compares these choices and highlights the need to preserve System 2 strengths while guiding efficient action execution.
- 1. Short Survey: Dual-system architectures differ in their selected modalities, policy training regimes, latent representations, and system-integration strategies.Table 1 distinguishes language, RGB, proprioception, depth, tactile, and point-cloud inputs, along with fine-tuning and scratch training.
- 1. Short Survey: VLAs use large multimodal models for visual-language understanding and generalization, while lightweight policies support efficient robotic action execution.The paper relates System 2 to computationally heavy MLLMs or VLAs and System 1 to efficient policy networks.
- 1. Short Survey: LCB uses LLaVA to produce a high-level <ACT> latent goal, which a pre-trained 3D Diffusion Actor combines with RGB and point-cloud inputs to generate actions.System 2 is fine-tuned with LoRA, while System 1 is fine-tuned conventionally.
- 1. Short Survey: DP-VLA uses OpenVLA representations from language and RGB observations to guide a Transformer policy that maps RGB and proprioception to actions.Its System 2 remains frozen while System 1 is trained from scratch.
- 1. Short Survey: HiRT and Robodual use different latent and policy designs, including MAP-pooled features, multimodal inputs, Perceiver Resampling, and DiT-based action generation.Robodual conditions action generation on task and action latents alongside distilled RGB, depth, tactile, and proprioceptive features.
- 1. Short Survey: The central design challenge is structuring information flow from the slower system to the faster system without losing generalization or execution efficiency.The paper identifies this balance as essential for high-performing and generalizable robotic systems.
1. MLLM Selection.
MLLM selection must balance robotic capability with inference cost and burden, while the value of robotic-data pretraining remains unresolved. The paper identifies domain-gap reduction and stronger language-instruction following as potential benefits of robotic pretraining.
- 1. MLLM Selection.: MLLM selection for robotics should balance sufficient task capability against lightweight inference and computational burden.Flower emphasizes spatial awareness and low-level vision, whereas MiniVLA uses Qwen-VL 0.25B to reduce inference costs.
- 1. MLLM Selection.: Whether robotic-data pretraining is necessary remains unresolved, although it may reduce domain gaps and improve language-instruction following.Robodual experiments are cited as evidence of especially robust instruction following after robotic-data exposure.
2. Policy Selection.
Small policy-model selection is relatively less controversial because DiT- and Flow Matching-based models are viewed as meeting current needs. New policy architectures and modality requirements remain open questions.
- 2. Policy Selection.: Current consensus holds that DiT- and Flow Matching-based small models can meet present robotic-policy needs.The paper treats this selection as relatively less controversial than other dual-system design choices.
- 2. Policy Selection.: Emerging architectures such as CARP and Dense Policy may introduce new downstream small-model designs and requirements.The paper also asks whether System 1 needs additional modalities and which information is essential.
3. Latent Feature Representation Selection.
Latent-feature selection is presented as the most complex and urgent dual-system research problem. Existing methods vary from single-layer embeddings to pooled multimodal features and learned action tokens, leaving downstream representations open for further study.
- 3. Latent Feature Representation Selection.: Latent-feature selection is the most complex dual-system design issue and urgently requires further research.The paper compares both dual-system and single-system approaches when discussing this problem.
- 3. Latent Feature Representation Selection.: Existing methods select latent information from different MLLM layers or pool language and visual features for downstream policy conditioning.Examples include DP-VLA's last-layer embedding, GR00T-N1's middle-layer features, and Roboflamingo and HiRT's max-pooled features.
- 3. Latent Feature Representation Selection.: Learned <ACT> tokens provide another bridge between upstream and downstream models, with later work extending the approach to multiple tokens and language features.LCB introduced the special-token approach, while Robodual further developed it with multiple <ACT> tokens.
- 3. Latent Feature Representation Selection.: Future work should explore more suitable latent representations for downstream action-generation models.The paper frames latent representation design as an important research focus for dual-system models.
4. MLLM Training Strategy.
MLLM training must preserve generalization while integrating effectively with downstream tasks. Existing approaches use freezing or fine-tuning, leaving improved fine-tuning techniques as an open research direction.
- MLLM training should preserve the model’s generalization capabilities without sacrificing downstream-task integration.
- Current MLLM training approaches primarily use frozen or fine-tuning methods.
- Developing better fine-tuning techniques remains a valuable research direction.
5. Policy Training Strategy.
Policy training focuses on reducing training cost, especially by fine-tuning pre-trained policies instead of training from scratch. Training from scratch also raises unresolved questions about optimization objectives and convergence.
- Policy training should reduce cost by considering fine-tuning from a pre-trained policy.Such fine-tuning could greatly reduce overall training time.
- Fine-tuning a pre-trained policy could greatly reduce the overall training time.
- Training policies from scratch leaves open whether different optimization objectives could make convergence difficult.
6. Dual-System Integration Strategy.
Dual-system integration embeds upstream latent information as a condition for downstream models. The strategy must address representation alignment and projector training, especially when the downstream policy is pre-trained.
- Dual-system integration embeds latent information as a condition into downstream models.
- Using CLIP loss to align upstream latent features with original text embeddings can limit downstream handling to trained cases.
- Dimension differences between upstream and downstream models commonly require an intermediate projector.
- When the downstream policy is pre-trained, the projector should be pre-aligned before training the MLLM.
- Simultaneously unfreezing and training the projector and MLLM can cause model training to collapse.
7. Dual-System Asynchronous Strategy.
Dual-system models use asynchronous strategies, including synchronous training with asynchronous testing. Inference-frequency differences may affect performance, but their impact depends on the effectiveness of the upstream features.
- LCB, HiRT, and Robodual employ different asynchronous strategies for dual-system models.
- LCB uses synchronous training but asynchronous testing as its asynchronous approach.The passage characterizes this as the most naive strategy.
- Differences in upstream and downstream inference frequency could affect final performance.
- The effect of asynchronous inference depends on whether the upstream features are effective in the first place.
2. Empirical Evaluations
The empirical evaluation standardizes key model components and compares policy training, MLLM training, integration, and asynchronous strategies across CALVIN settings. Results emphasize dynamic robustness, pre-trained policy fine-tuning, prompt tuning, projector pre-alignment, and latent behavior.
- 2.1. Experiment setup: The study fixes the MLLM, downstream policy, latent integration, and asynchronous training/testing settings to isolate selected design choices.LLaVA1.0, 3DDA, LCB-style integration, and synchronous training with asynchronous testing are used for consistency.
- 2.1. Experiment setup: CALVIN-D evaluates robustness using static objects and four movement patterns, while CALVIN-E tests language-instruction generalization.The dynamic evaluation uses Left, Forward, Diagonal, and Circle object movements.
- 2.3.1. Policy Training Strategy: Pre-trained policy fine-tuning improves performance while reducing training time, so subsequent experiments use this paradigm.The comparison holds the large-model configuration constant and changes whether the downstream policy is pre-trained or trained from scratch.
- 2.3.2. MLLM Training Strategy: Prompt tuning preserves generalization better than fine-tuning or freezing in language-generalization experiments, especially with CLIP-loss supervision.Prompt tuning trains a new <ACT> token through the lm-head while keeping other MLLM parameters fixed.
- 2.3.3. Dual-System Integration Strategy: Projector pre-alignment is important because direct connection of upstream and downstream models fails under frozen, fine-tuning, and prompt-tuning approaches.The comparison contrasts joint training with a separate stage that first trains the projector and downstream policy.
- 2.4. Testing strategy of dual system: The latent embedding primarily encodes target-object, spatial-relation, and action semantics, while action-token probabilities can remain stable across time.The reported word-similarity analysis also observes noise and consistently higher probability for “right” than “left.”
3. A Simple yet Effective Dual System VLA
The proposed dual-system VLA uses a frozen MLLM with a learnable prompt token to produce latent guidance for a downstream policy, while auxiliary action prediction encourages multimodal reasoning. Experiments examine this design with staged training and report gains from integrating the two layers and adding auxiliary tasks, while asynchronous inference has little effect.
- Architecture: The system combines a pre-trained MLLM and policy, with the MLLM producing a latent embedding that conditions the low-level policy’s action prediction.The policy predicts action trajectories and gripper states over a temporal horizon from observations, latent guidance, and proprioception.
- Prompt Tuning: A learnable <ACT> token is appended to the instruction while all MLLM parameters remain frozen, adapting the model through prompt tuning.The new token embedding is the only MLLM parameter updated during prompt tuning.
- Multimodal Reasoning Learning: An auxiliary task predicts action trajectories and gripper actions from the prompt-token embedding, requiring the latent representation to combine visual and textual information.The auxiliary objective uses L1 loss for 3D locations and rotations and binary cross-entropy for gripper opening.
- Training: Training proceeds in two stages: first aligning prompt and projection layers with frozen models, then fine-tuning the low-level policy while keeping the large model frozen.The first stage performs pre-alignment; the second jointly trains the policy with the prompt and projection components.
- Results: The final evaluation on CALVIN ABC-D reports that integrating upper and lower layers improves language generalization, auxiliary tasks improve standard and generalization performance, and asynchronous inference has little impact.The table reports success rates and average task completion length across five tasks; Asy(10) denotes a 10-step delay and AUX denotes auxiliary tasks.
4. Discussion & Limitation
The project remains an initial step toward a full open-source reproduction of Helix, with several deployment and collaboration goals still unmet. The authors plan continued updates while acknowledging that some claims remain insufficiently verified.
- The authors state that a full open-source reproduction of Helix remains incomplete.
- Unmet goals include real-robot deployment, faster downstream policy execution, physical-robot operation, humanoid deployment, and humanoid-robot collaboration.
- The report is described as an initial version, with continued updates intended to address the listed open-source objectives.
- The authors acknowledge that some claims in the article have not yet been fully verified.