Source-linked AI summary
Hume: Introducing System-2 Thinking in Visual-Language-Action Model
Haoming Song, Delin Qu, Yuanqi Yao, Qizhi Chen, Qi Lv, Yiwen Tang, Modi Shi, Guanghui Ren, Maoqing Yao, Bin Zhao, Dong Wang, Xuelong Li
TL;DR
Hume addresses the limited exploration of slow, System-2 thinking for robotic foundation models performing complex physical-world tasks. It combines value-guided candidate selection with cascaded action denoising in a dual-system VLA model, outperforming existing state-of-the-art VLA models across simulation benchmarks and real-robot deployments.
Problem
Slow, reflective System-2 thinking has improved LLMs in digital domains but remains largely unexplored for robotic foundation models interacting with the physical world.
Method
Hume combines low-frequency value-guided System-2 candidate selection with a high-frequency reactive System-1 policy that performs cascaded action denoising.
Results
Hume outperforms existing state-of-the-art VLA models across simulation benchmarks and real-robot deployments, including a +4.4% success-rate increase over π0 on LIBERO.
Takeaways & Limitations
The results support Hume as a promising direction for generalist robot policies requiring rapid, fluid control of complex dexterous tasks.
Takeaways & Limitations
Value-guided System-2 thinking is limited by the quality of sampled candidate action chunks, and its estimated state-action values are not well-aligned with semantics.
Abstract
from arXiv · showhide
Humans practice slow thinking before performing actual actions when handling complex tasks in the physical world. This thinking paradigm, recently, has achieved remarkable advancement in boosting Large Language Models (LLMs) to solve complex tasks in digital domains. However, the potential of slow thinking remains largely unexplored for robotic foundation models interacting with the physical world. In this work, we propose Hume: a dual-system Vision-Language-Action (VLA) model with value-guided System-2 thinking and cascaded action denoising, exploring human-like thinking capabilities of Vision-Language-Action models for dexterous robot control. System 2 of Hume implements value-Guided thinking by extending a Vision-Language-Action Model backbone with a novel value-query head to estimate the state-action value of predicted actions. The value-guided thinking is conducted by repeat sampling multiple action candidates and selecting one according to state-action value. System 1 of Hume is a lightweight reactive visuomotor policy that takes System 2 selected action and performs cascaded action denoising for dexterous robot control. At deployment time, System 2 performs value-guided thinking at a low frequency while System 1 asynchronously receives the System 2 selected action candidate and predicts fluid actions in real time. We show that Hume outperforms the existing state-of-the-art Vision-Language-Action models across multiple simulation benchmark and real-robot deployments.
1 Introduction
Hume addresses the limited exploration of System-2 slow thinking for robotic foundation models, where physical-world control requires both deliberation and real-time responsiveness. It introduces a dual-system VLA model combining value-guided candidate selection with cascaded action denoising, and reports gains across simulation and real-robot evaluations.
- Motivation: System-2 thinking remains largely unexplored for robotic foundation models despite its relevance to complex physical-world tasks.The paper contrasts deliberate System-2 thinking with fast System-1 intuition and identifies robotic deployment as an open problem.
- Challenges: Robotic System-2 thinking faces semantic ambiguity in delicate actions and a tension between slow reasoning and real-time control.These challenges complicate direct transfer of text-based reasoning techniques to dexterous robot action prediction.
- Approach: Hume is a dual-system VLA model that equips System 2 with value-guided repeat sampling and combines it with cascaded action denoising.The design is intended to combine low-frequency deliberation with high-frequency control.
- Results: The paper reports state-of-the-art performance across multiple benchmarks and real-robot tests, including challenging humanoid control tasks.The authors present Hume as evidence that human-like thinking capabilities can be explored in generalist robot policies.
- Results: +4.4% success rate over π0 on LIBERO, +25.9% on SimplerEnv, and +12.9% improvement in real-world deployments are reported for Hume.These figures are presented as headline performance gains across simulation benchmarks and real-robot tests.
2 Related Work
Related work connects Hume to VLA models for robot control, dual-system architectures for efficient interaction, System-2 reasoning methods, and cascaded denoising techniques.
- Dual-System Vision-Language-Action Models: Prior VLA models extend vision-language models to robot control using discretized or continuous action representations.Examples include RT-2, OpenVLA, and π0, which use action tokens, adapted VLMs, or flow matching.
- Dual-System Vision-Language-Action Models: Dual-system robot architectures run slower high-level models alongside faster control policies to support real-time interaction.HiRT is cited as an example of low-frequency VLM processing paired with high-frequency vision-based control.
- System 2 and System-2 Thinking: System-2 reasoning methods generate intermediate reasoning, explore multiple solution paths, reflect on previous attempts, or use self-verification.The related methods include Chain-of-Thought, Tree-of-Thoughts, and Reflexion.
- Cascaded Denoising: Cascaded denoising has been used in diffusion models for progressive image generation and signal restoration.Prior approaches cascade resolution, function-learning, noise-prediction, image-prediction, or spectral reconstruction stages.
3 Methodology
Hume uses System 2 to generate and value-rank partially denoised action candidates, then uses a faster System 1 to refine them asynchronously into continuous robot actions. Its value estimator is trained with offline reinforcement learning and calibrated Q-learning, while the overall model uses staged training.
- Methodology: Hume organizes methodology around value-guided System-2 thinking, asynchronous dual-system cooperation, and multi-stage training and deployment.The architecture and strategy are described across value estimation, cascaded denoising, and training sections.
- 3.1 Value-Guided System-2 Thinking: System 2 generates N action chunks with different noise levels and selects the candidate with the highest estimated state-action value.The value-query head ranks candidate actions conditioned on the observation and action representation.
- 3.1 Value-Guided System-2 Thinking: Candidate actions model p(A_t|o_t) through transformer-based flow matching that gradually removes noise from randomly initialized actions.The observation includes images, language instructions, and robot state; τ represents the flow-matching noise level.
- 3.1 Value-Guided System-2 Thinking: The denoising head uses 10 denoising steps with δ = 0.1 during training.Training minimizes the discrepancy between the actual remaining noise and the network’s predicted vector field.
- 3.1 Value-Guided System-2 Thinking: The value-query head estimates Qθ(q_t,A_t) using a special query token and is trained on demonstrations with offline reinforcement learning and calibrated Q-learning.The training reward assigns +1 to the last three transitions of an episode and 0 elsewhere.
- 3.2 Cascaded Dual-system Action Denoising: System 1 receives the selected candidate, segments it into K := H/h sub-action chunks, and continuously denoises them using current observations.System 1 is lightweight and faster than System 2, allowing it to finish refinement before the next selected action chunk arrives.
4 Experiment
Hume is evaluated across diverse simulation and real-world robot-control settings, including long-horizon, dexterous, and humanoid tasks. It achieves strong performance across benchmarks and platforms, while ablations show that cascaded denoising, value-guided selection, and repeat sampling are important design components.
- Evaluation Setup: Experiments span 3 simulation environments and 3 real-world robotic platforms, covering 15 robot learning scenarios and 21 real-world manipulation tasks.The evaluation includes WidowX, AgiBot G-1, and Franka robots.
- Simulation Benchmarks: Hume achieves the highest average LIBERO success rate at 98.6% and ranks first across the evaluated policies.On LIBERO-Long, it reaches 96.7% success rate, with gains of +11.5% over π0 and +6.1% over GR00T.
- Simulation Benchmarks: On SimplerEnv, Hume reaches 72.6% average success on WidowX and 76.4% on Google robot tasks, exceeding π0 by +32.5% and +19.6%, respectively.The benchmark varies lighting, colors, textures, and camera poses to test robustness across visual conditions.
- Real-World Evaluation: Hume attains 91% average success across complex unseen WidowX tasks, improving by +12% over π0 and +33% over OpenVLA.The reported advantage is associated with selecting alternative trajectories after initial failures.
- Real-World Evaluation: On Franka and Agibot G-1 tasks, Hume achieves 82% success on Pour Water and 87% average success across Franka tasks.The reported improvements are +20% over π0 and +60% over GR00T for Pour Water, and +14.75% over π0 and +37.25% over OpenVLA on Franka.
- Ablations on Design Decisions: Removing cascaded denoising reduces performance by -3.2% on SimplerEnv, -2.7% on LIBERO, and -19% in real-world tasks, while removing System 1 causes larger decreases of -9.8%, -8.8%, and -63%.The ablations attribute these drops to reduced candidate diversity and the inability to remove remaining action noise for precise movements.
- Ablations on Design Decisions: Removing value-guided thinking causes average declines of -14.95% on SimplerEnv, -13.7% on LIBERO, and -78% in real-world tasks; removing repeat sampling causes declines of -6.2%, -4.8%, and -37%.These variants respectively use random candidate selection or generate only one candidate, limiting useful action selection.
5 Conclusion and Limitations
Hume combines value-guided System-2 thinking with System-1 cascaded action denoising for dexterous robot control and reports strong performance across robot tasks. The authors identify limitations in candidate sampling, value semantics, and the simplicity of the current thinking paradigm.
- Hume integrates value-guided System-2 thinking with cascaded action denoising to support rapid, fluid control for dexterous robot tasks.
- Extensive simulation and real-robot experiments report that Hume outperforms current state-of-the-art models across various robot tasks.
- Limitations: Hume’s value-guided System-2 thinking is limited by the quality of sampled candidate action chunks.
- Limitations: The estimated state-action values are not well-aligned with semantics, motivating further research on value learning.
- Limitations: Hume’s System-2 thinking remains relatively naive compared with potential tree search, self-correction, and reinforcement-learning approaches.
A Hume Visualization Analysis
This section introduces visual analyses of Hume’s inference process and its two key designs: value-guided System-2 thinking and cascaded action denoising.
- The appendix demonstrates Hume’s detailed inference process through the dexterous Push-T task.
- The appendix visualizes value-guided System-2 thinking and cascaded action denoising to provide a comprehensive understanding of Hume.
A.1 Hume Workflow Visualization
Hume samples multiple candidate actions, selects one using estimated state-action values, and continuously denoises it into smoother executed actions. The visualizations cover value maps, temporal inference, and action dimensions.
- Hume Workflow Visualization: At Push-T inference steps t = 0, 30, 60, 90 with horizon H = 30, System 2 produces 10 candidate actions per timestep.
- Hume Workflow Visualization: System 2 selects the highest-value candidate, which System 1 continuously denoises into the executed action.
- Hume Workflow Visualization: The final System-1 denoised action is smoother and more delicate for accomplishing the Push-T task.
- Value-Guided System-2 Thinking: In LIBERO-GOAL, candidate actions are projected from the 7-dimensional action space onto two dimensions with PCA and colored by estimated state-action value.
- Value-Guided System-2 Thinking: Ground-truth actions consistently occupy high-value regions, while not always appearing at the highest-value positions in the value map.
- Cascaded Action Denoising: The LIBERO-OBJECT visualization groups the seven action dimensions into X-Y, X-Z, Y-Z, and R-P coordinate pairs.
B Implementation Details
Hume’s value-query head is trained to estimate state-action values using Bellman-error minimization with a conservative regularizer for out-of-distribution actions.
- Value Objective Functions: The value-query head learns Qθ(qt, At), an estimate of the state-action value function in a Markov Decision Process.
- Value Objective Functions: Its training objective minimizes Bellman error together with a regularization term R(θ).
- Value Objective Functions: The discount factor γ lies in (0, 1), while S and A denote the state and action spaces and P and r denote dynamics and reward functions.
- Value Objective Functions: The second term is a standard temporal-difference error using the value-query output and a backup operator applied to a delayed target network.
- Value Objective Functions: The calibrated conservative regularizer penalizes Q-values for out-of-distribution actions and compensates for this pessimism on actions observed in the training dataset.
B.2 Training and Inference Hyperparameters
Hume uses environment-specific training and inference settings across simulation and three real-world robot platforms.
- LIBERO uses third-person and wrist cameras plus robot state, with System 2 and System 1 chunk sizes of 16 and 8.
- SimplerEnv uses third-person images and robot state, with Bridge chunk sizes of 8 and 4 and Google Robot sizes of 4 and 2.
- Franka uses third-person images and robot state, with System 2 and System 1 chunk sizes of 16 and 8.
- WidowX uses Bridge-like settings, with third-person images, robot state, and System 2 and System 1 chunk sizes of 8 and 4.
- Agibot G-1 uses head and bilateral wrist cameras plus robot state, with System 2 and System 1 chunk sizes of 30 and 15.
C Experiment Details
The experiment-details section directs readers to appendices for evaluation setups, comparison-method implementations, testing standards, and detailed results.
- Appendix C.1 provides simulation benchmark setups, comparison-method implementation details, and detailed test results.
- Appendix C.2 provides real-robot task setups, testing standards, and detailed test results.
C.1 Simulation Benchmark Details
Simulation evaluations use established camera and robot-state inputs while combining reported or reproduced results across several policy baselines.
- LIBERO methods use third-person and wrist cameras plus robot state as input.
- LIBERO results combine technical-report results, repository-provided results, and results obtained by training and testing open-source code.
- SimplerEnv includes detailed Google Robot results across varying environment layouts, object positions, and textures.
C.2 Real-World Evaluation Details
Real-world evaluation spans WidowX 250s, Franka-Emika-Panda, and AgiBot G-1, covering pick-and-place, articulated, instruction-following, long-horizon, and deformable-object tasks.
- Evaluation platforms: Evaluations cover three real-world platforms: WidowX 250s, Franka-Emika-Panda, and AgiBot G-1.
- WidowX 250s: WidowX tasks include object discrimination, spatial awareness, articulated-door manipulation, localization, and varied-height cup placement.
- Franka-Emika-Panda: Franka tasks include limited-data banana transfer, bowl placement, revolute-joint manipulation, spatial variation, and color-specific cube placement.
- AgiBot G-1: AgiBot tasks test snack-restocking generalization, long-horizon water pouring, two-arm shorts folding, and human-collaborative bottle passing.
- Evaluation metrics: Real-world evaluation reports overall success for simple tasks, partial and overall success for general tasks, and subtask plus overall success for long-horizon tasks.
- Generalization tests: Tests vary lighting, object types, environmental layouts, language instructions, and shorts colors and materials to assess generalization.
- Failure recovery: Hume more often recovers from failure states by sampling candidate actions and selecting one using estimated state-action value.