Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

SwitchVLA: Execution-Aware Task Switching for Vision-Language-Action Models

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SwitchVLA lets one robot policy roll back or advance when the user changes the instruction mid-motion.

desk verdict A genuinely new training-time way to make VLA policies switch tasks, with strong but not airtight results; it needs a pseudocode/text fix, a neutral success metric, and released code before the details can be trusted. read the letter →

arxiv 2506.03574 v1 pith:AVNCDBOZ submitted 2025-06-04 cs.RO

Meng Li , Zhen Zhao , Zhengping Che , Fei Liao , Kun Wu , Zhiyuan Xu , Pei Ren , Zhao Jin
show 2 more authors
Ning Liu Jian Tang
This is my paper
classification cs.RO
keywords ImitationLearningVision-Language-ActionTaskSwitchingRobotManipulationContactStateBehaviorModulationFlowMatchingLIBERO
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a vision-language-action (VLA) robot policy can handle mid-execution instruction changes by treating task switching as behavior modulation instead of re-planning. The key idea is to condition the policy on a binary contact state and a three-valued behavior mode—forward, rollback, advance—so that when a new instruction arrives, the model decides whether to keep going, undo its recent actions, or move toward the new task's starting pose. No external planner and no extra switch-specific demonstrations should be needed, because the required labels come from segmenting ordinary expert trajectories into contact phases. If this holds, robots in retail or household settings could respond to changed intentions the way a human assistant would, setting down the old object before taking up the new one. Simulation and two real Franka workstations show large gains over standard VLA baselines, especially when the switch happens while the gripper is already in contact.

What carries the argument

The central object is the Conditional Execution Expert, a Florence-2-based transformer that jointly predicts the binary contact state $c_t$, the behavior mode $b_t \in \{0: \text{forward}, 1: \text{rollback}, 2: \text{advance}\}$, and a $K$-step action chunk from visual tokens plus previous and current instruction tokens and the previous contact state. The load-bearing mechanism is the training data-sampling strategy (Algorithm 1), which converts each expert trajectory into forward, rollback, and advance samples using instruction mismatch and contact state as the switch signal, with the advance target $f(a_t, l_{\text{cur}})$ defined as linear joint-space interpolation to the mean start pose of the new task's demonstrations. A flow-matching loss on action chunks plus classification losses on contact and behavior mode train the policy, and temporal ensembling over the first 30 actions of each chunk smooths execution.

What would settle it

Take a mid-switch pair where the mean start pose of the new task lies inside an obstacle or beyond the robot's reach; if SwitchVLA's advance behavior cannot complete the new task in that configuration, the start-pose assumption fails. A second check is to corrupt or remove the contact labels at evaluation time; if mid-switch success stays high, the contact signal is not doing the claimed work.

Watch

Extended reading notes

Core claim

SwitchVLA's central claim is that one VLA policy can execute both single-task demonstrations and task switches, as long as it is trained to jointly predict a contact state, a behavior mode, and an action chunk at every step. During training, ordinary single-task episodes are re-sampled into three kinds of examples: forward uses the original action chunk with the matching instruction; rollback reverses the previous K actions when the new instruction mismatches and the gripper is in contact; advance linearly interpolates in joint space from the current pose toward the mean of the new instruction's initial trajectory poses. At inference, the policy receives the previous and current instructions plus the previous contact state, predicts which mode applies, and either continues, undoes, or moves toward the next task. The paper reports that this recipe produces smooth, reactive switching without external planners or switch-specific data, and outperforms π0 and OpenVLA-OFT in LIBERO-Goal simulation and MT-ACT, Diffusion Policy, and π0 on two dual-arm Franka workstations.

Load-bearing premise

The load-bearing premise is that a binary contact state and three discrete behavior modes faithfully represent task progress and switching intent, and that every candidate next task has a usable 'start pose' computed as the mean of its earlier demonstrations, so the advance behavior can interpolate toward it without any new switch data.

Editorial extensions

If this is right

  • Mid-grasp instruction changes become recoverable: the rollback behavior can set an object down cleanly before the new task starts, instead of oscillating or dropping it.
  • Switching decisions are made inside the policy, so the robot can react at control frequency without waiting for a high-level planner's reply.
  • The advance behavior removes the assumption that every task starts from a fixed home pose, letting the robot travel to the next task's own starting configuration.
  • Long chains of task switches can be composed from single-task data alone, with the paper reporting accumulated success through six consecutive switched tasks in simulation and five in the real world.
  • Because contact and behavior mode are auxiliary prediction heads, the same backbone can keep its single-task performance while gaining switch capability, with no architecture change at deployment.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: the binary contact state is a progress proxy; tasks whose meaningful phases never involve gripper-object contact would likely need a different execution signal, which the paper's limitations section acknowledges.
  • Editorial extension: the advance behavior depends on a per-task canonical start pose computed from that task's old demonstrations, so the 'no additional data' claim applies to tasks already seen in training; a genuinely new instruction would need a synthesized start pose.
  • Editorial extension: one could test whether the model truly uses contact to infer progress by clamping the predicted contact state to a constant during a mid-switch evaluation; if success holds, the progress signal is less load-bearing than claimed.
  • Editorial extension: the three discrete modes compress intent changes into a single label; partial or iterative corrections would require a continuous behavior representation, a direction the authors explicitly leave open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes SwitchVLA, a vision-language-action (VLA) model that supports mid-execution task switching by conditioning action generation on a binary contact state and a three-valued behavior mode (forward, rollback, advance). Training uses only single-task demonstrations; rollback targets are synthesized by reversing action chunks, and advance targets are produced by linear interpolation in joint space from the current pose toward a per-task canonical start pose defined as the mean of initial trajectories for the new instruction. The method is evaluated in LIBERO-Goal simulation and on two real Franka workstations against π0, OpenVLA-OFT, MT-ACT, and Diffusion Policy, reporting large improvements in switch success rates and comparable single-task performance.

Significance. If the central claim holds, the paper makes a useful step toward interactive VLA control: it demonstrates that a single policy, trained on ordinary single-task demonstrations, can produce coherent rollback/advance behavior at inference, and the reported gains over strong baselines are large and replicated across simulation and two real platforms. The behavioral decomposition into forward/rollback/advance is simple and interpretable. However, two load-bearing issues prevent immediate acceptance: (i) the 'no additional switch-specific data' claim is undermined by the per-task canonical start poses required for advance, and (ii) the evaluation success criterion embeds the paper's own behavior specification, which may bias comparisons against baselines that solve the task differently. The paper would be substantially strengthened by releasing code/data, correcting the Algorithm 1 vs. text inconsistency, and adding a task-completion-only evaluation metric.

major comments (3)
  1. [Algorithm 1; Sections 3.2 and 3.4.1] Algorithm 1 lines 12-17 assign rollback when c_pre = c_0 (no contact) and advance otherwise, which is the opposite of the definitions in Section 3.2 ('rollback ... upon detecting an intent mismatch while in contact' and 'advance ... when the instruction updates and no physical interaction is present') and of the training bullets in Section 3.4.1 (rollback requires active contact c_i_t = 1; advance assumes no contact). This makes the actual convention used by the trained policy ambiguous, and a reader cannot tell which condition was used to generate the behavior-mode labels. The authors must correct this inconsistency and state explicitly which convention was implemented, ideally with per-mode classification accuracy to confirm the model learned the intended mapping.
  2. [Section 3.4.1; Section 4.2] The advance behavior, which is the mechanism responsible for mid- and late-phase switching performance, relies on a per-task canonical start pose a_normal_0 defined as the mean of initial trajectories for instruction l_cur, and Algorithm 1 line 16 computes f(a_t, l_cur) as linear interpolation toward that pose. This pose must be computed from demonstrations of every task the robot can switch to, so the claim 'without additional switch-specific data' in the Abstract and Section 1 is not supported as stated: for a genuinely new task l_cur, no initial trajectories exist. In the experiments all switch targets come from the fixed training task set, so the general claim of switching to arbitrary new tasks is not demonstrated. The authors should clarify this data requirement, release the canonical poses and interpolation schedule (or code), and provide an ablation that isolates the contribution of the hand-designed advance interpolation from the learned policy components.
  3. [Section 4.1; Tables 1-3] The success criterion in Section 4.1 requires that after a new instruction is issued, 'Task A behaves as expected formulated in Section 3' before Task B is completed. Since Section 3 defines the expected behavior as the rollback/advance protocol, this criterion embeds SwitchVLA's own behavior specification into the evaluation. Baselines that complete both tasks without following the rollback/advance protocol—for example, by immediately executing Task B from the current state or by finishing Task A first—are scored as failures even if both task goals are met. The headline margins in Tables 1-3 may therefore overstate the advantage in raw task completion. The authors should report an additional evaluation that counts a trial as successful whenever both instructions are ultimately satisfied, regardless of intermediate behavior, and should break down baseline failures into those caused by protocol non-compliance versus genuine task failure.
minor comments (6)
  1. [Section 4.1] The phrase 'behaves as expected formulated in Section 3' should cite the specific definitions (e.g., the behavior bullets in Section 3.4.1) rather than the entire section, as the current reference is too broad for a reader to verify the evaluation protocol.
  2. [Section D.3] The sentence stating that baselines 'by chance fail at the beginning S2→S5 pairwise tasks' is imprecise, since the evaluation is deterministic; please describe the observed failure mode and stage instead of attributing it to chance.
  3. [Table 2 note] The note 'while individual tasks are not restricted to the same one, we ensure the use of distinct task pairs appeared in the long sequence' is confusing given that the sequences in Table 8 contain repeated tasks (e.g., LR1: R1.1→R1.3→R1.2→R1.1→R1.2→R1.4); please clarify what is meant by 'distinct task pairs' and how repetition is handled.
  4. [Figure 5 caption] The caption contains a duplicated conjunction: 'smooth and consistent and instruction-aligned' should read 'smooth, consistent, and instruction-aligned.'
  5. [Table 4] The contact state labels have only 80.7% average accuracy; since the contact state is the sole execution-feedback signal, the paper would benefit from a discussion of how label noise affects behavior-mode classification and downstream switching performance.
  6. [General] The paper does not include a reproducibility statement or a link to released code/data; given that the advance mechanism depends on per-task canonical poses, at least the pose-extraction code and interpolation parameters should be released to allow verification.

Circularity Check

1 steps flagged · score 3.0 of 10

Switching success is defined partly by the paper's own rollback/advance protocol; final Task B completion keeps the result partially independent.

  1. self definitional [Section 4.1, Task Protocol]
    "A task switching is considered successful only if Task A enters its designated execution phase without failure, and when new instruction is triggered, Task A behaves as expected formulated in Section 3, and then Task B is subsequently completed."

    The success criterion for every switch trial includes 'Task A behaves as expected formulated in Section 3.' Section 3 formulates the expected behavior as the paper's own three-mode scheme, in particular rollback under contact and advance when no contact. Thus a baseline that receives a new instruction and directly completes Task B, without first executing the paper's rollback/advance protocol, is scored as a failure no matter whether Task B succeeds. The reported large margins over pi0, OpenVLA-OFT, DP, and MT-ACT in mid/late switching are therefore partly built into the metric definition. The additional requirement that Task B be subsequently completed is externally grounded, so the circularity is partial rather than total.

full rationale

SwitchVLA is an empirical policy paper, not a formal derivation, and most of its components are not circular in a mathematical sense. There is no load-bearing self-citation chain, and no fitted parameter is renamed as a prediction: contact and behavior labels are weak supervision derived from trajectories, rollback and advance targets are constructed from existing task demonstrations, and the appendix's own additional-data comparison explicitly concedes that a baseline with enough curated switching data could match SwitchVLA, which undercuts any claim that the comparison is forced by construction. The main circularity risk is the evaluation protocol: Section 4.1 defines a successful switch partly as 'Task A behaves as expected formulated in Section 3,' meaning the metric encodes the paper's own rollback/advance behavior as a necessary condition. Because Task B completion is still required and externally verifiable, the circularity is partial. Algorithm 1's reversed contact condition for rollback/advance relative to Section 3.2 is a consistency defect, not a circularity, and the per-task canonical start pose is a data-efficiency caveat rather than an equivalence of input and output.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The method's central claims rest on: (1) per-task canonical start poses fitted from demonstrations for the advance behavior; (2) a binary contact state serving as an adequate proxy for task progress; (3) the sufficiency of three discrete behavior modes; and (4) the availability of both previous and current instructions at inference. These are modeling choices and domain assumptions, not derived facts. No new physical entities are proposed; the behavior-mode variable is an invented latent construct with no independent falsifiable handle outside the paper's experiments.

free parameters (4)
  • per-task canonical start pose = mean of initial expert trajectories for l_cur
    Computed from training demonstrations; needed by the advance behavior, which linearly interpolates to this pose.
  • contact state sampling interval = 1/3 Hz
    GPT-4o labeling samples at 1/3 Hz with sliding-window mode filtering, a hand-chosen parameter affecting label noise.
  • loss weights for auxiliary heads = 1e-2 (contact BCE), 1e-4 (behavior CE)
    Hand-chosen in Section B.2; the balance of supervision objectives affects contact/behavior accuracy and downstream switching performance.
  • current-task sampling probability P_cur = not specified
    Controls the mixture of forward vs. switch-conditioned training samples in Algorithm 1.
assumptions (5)
  • domain assumption Binary contact state is a sufficient proxy for task progress and switch timing.
    Stated in Limitations: 'it relies on contact-based phase segmentation as a proxy for task progress.' Non-contact transitions or abstract semantics are not modeled.
  • ad hoc to paper Three discrete behavior modes (forward, rollback, advance) suffice to represent all relevant task-switching behaviors.
    The paper itself notes ambiguous or drifting intent is not handled; the mode set is a designer choice, not derived.
  • domain assumption Both previous and current instructions are available at inference, and the model can detect when the instruction changes.
    Algorithm 1 and Section 3.4.2 condition on (l_pre, l_cur); the evaluation assumes the system knows the new instruction.
  • domain assumption Expert demonstrations can be labeled with contact states automatically at sufficient accuracy.
    Section B.3 uses GPT-4o with about 80.7% accuracy; if labels are poor, rollback/advance supervision degrades.
  • domain assumption A per-task canonical start pose is available for any possible new task.
    Advance behavior interpolates to the mean of initial trajectories of l_cur; for a novel task this must be computed from new demonstrations.
invented entities (1)
  • Behavior mode
    purpose: Latent three-valued condition (forward/rollback/advance) used to supervise and condition action generation during task switching.
    A model construct introduced by the paper; its validity is only demonstrated by the paper's own experiments, with no external benchmark or falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SwitchVLA: Execution-Aware Task Switching for Vision-Language-Action Models." pith.science (2026). https://pith.science/paper/AVNCDBOZ

@misc{pith2026250603574,
  author       = {Pith},
  title        = {Pith review of: SwitchVLA: Execution-Aware Task Switching for Vision-Language-Action Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AVNCDBOZ}},
  note         = {Machine review of arXiv:2506.03574}
}
read the original abstract

Robots deployed in dynamic environments must be able to not only follow diverse language instructions but flexibly adapt when user intent changes mid-execution. While recent Vision-Language-Action (VLA) models have advanced multi-task learning and instruction following, they typically assume static task intent, failing to respond when new instructions arrive during ongoing execution. This limitation hinders natural and robust interaction in dynamic settings, such as retail or household environments, where real-time intent changes are common. We propose SwitchVLA, a unified, execution-aware framework that enables smooth and reactive task switching without external planners or additional switch-specific data. We model task switching as a behavior modulation problem conditioned on execution state and instruction context. Expert demonstrations are segmented into temporally grounded contact phases, allowing the policy to infer task progress and adjust its behavior accordingly. A multi-behavior conditional policy is then trained to generate flexible action chunks under varying behavior modes through conditioned trajectory modeling. Experiments in both simulation and real-world robotic manipulation demonstrate that SwitchVLA enables robust instruction adherence, fluid task switching, and strong generalization-outperforming prior VLA baselines in both task success rate and interaction naturalness.

Figures

Figures reproduced from arXiv: 2506.03574 by the authors.

Figure 1
Figure 1. (a) Processes 1 and 2 show normal task execution. When the user changes their mind (e.g., “pick up lemon place on plate”), conventional VLA models cannot adjust its plan, leading to erratic behavior like oscillation or dropping items, as seen in Process 3. (b) A more natural response involves returning the previously held item (e.g., placing down the cookie box in Process 3 and then picking up the lemon and placing … view at source ↗
Figure 2
Figure 2. Overview of SwitchVLA. The framework consists of the Vision-Language-Contact Em [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Identify and label time intervals of a specified event from trajectory data using a pre-trained [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Illustration of the training pipelines for [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Top: Performance of π0 [9] under pairwise task switching. (a), (b), and (c) each illustrate a unique task transition. Sudden switches during execution lead to erratic behaviors. Bottom: SwitchVLA enables smooth and consistent and instruction-aligned task transitions. E…
Figure 6
Figure 6. Figure 6: The proportion of failure causes during pairwise task switching in real-world experiments. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: (a) Dual-armed Franka workstation 1 setup. (b) Dual-armed Franka workstation 2 setup. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. XR-1: Towards Versatile Vision-Language-Action Models via Learning Unified Vision-Motion Representations

    cs.RO 2025-11 unverdicted novelty 6.0 of 10

    XR-1 introduces Unified Vision-Motion Codes learned by dual-branch VQ-VAE and applies them in a three-stage training pipeline to outperform prior VLA models on 120+ real-world manipulation tasks across six robot embodiments.

  2. Robotic Manipulation via Imitation Learning: Taxonomy, Evolution, Benchmark, and Challenges

    cs.RO 2025-08 conditional novelty 4.0 of 10

    A survey that taxonomizes robotic manipulation policies trained by imitation learning, traces their evolution, and compiles benchmark comparisons.

Reference graph

Works this paper leans on

38 extracted references · 6 canonical work pages · cited by 2 Pith papers

  1. [1]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Haus- man, A. Herzog, J. Hsu, et al. Rt-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

    arXiv 2022

  2. [2]

    Bharadhwaj, J

    H. Bharadhwaj, J. Vakil, M. Sharma, A. Gupta, S. Tulsiani, and V . Kumar. Roboagent: Towards sample efficient robot manipulation with semantic augmentations and action chunking.arxiv, 2023

    work page 2023

  3. [3]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 2023

    work page 2023

  4. [4]

    Goyal, J

    A. Goyal, J. Xu, Y . Guo, V . Blukis, Y .-W. Chao, and D. Fox. Rvt: Robotic view transformer for 3d object manipulation.CoRL, 2023

    work page 2023

  5. [5]

    Zhang, Y

    J. Zhang, Y . Guo, X. Chen, Y .-J. Wang, Y . Hu, C. Shi, and J. Chen. Hirt: Enhancing robotic control with hierarchical robot transformers.arXiv preprint arXiv:2410.05273, 2024

    arXiv 2024

  6. [6]

    Z. Fu, T. Z. Zhao, and C. Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. InConference on Robot Learning (CoRL), 2024

    2024

  7. [7]

    Cheang, G

    C.-L. Cheang, G. Chen, Y . Jing, T. Kong, H. Li, Y . Li, Y . Liu, H. Wu, J. Xu, Y . Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation. arXiv preprint arXiv:2410.06158, 2024

    arXiv

  8. [8]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    arXiv 2024

Show all 38 references
  1. [9]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al.pi 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    2024 arXiv

  2. [10]

    Z. Wu, Y . Zhou, X. Xu, Z. Wang, and H. Yan. Momanipvla: Transferring vision-language- action models for general mobile manipulation.arXiv preprint arXiv:2503.13446, 2025

    2025 arXiv

  3. [11]

    M. J. Kim, C. Finn, and P. Liang. Fine-tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645, 2025. 9

    2025 arXiv

  4. [12]

    H. Yuan, C. Zhang, H. Wang, F. Xie, P. Cai, H. Dong, and Z. Lu. Plan4MC: Skill reinforcement learning and planning for open-world Minecraft tasks.arXiv preprint arXiv:2303.16563, 2023

    2023 arXiv

  5. [13]

    Namasivayam, H

    K. Namasivayam, H. Singh, V . Bindal, A. Tuli, V . Agrawal, R. Jain, P. Singla, and R. Paul. Learning neuro-symbolic programs for language guided robot manipulation. In2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023

    2023

  6. [14]

    Huang, C

    W. Huang, C. Wang, Y . Li, R. Zhang, and L. Fei-Fei. Rekep: Spatio-temporal reasoning of relational keypoint constraints for robotic manipulation.arXiv preprint arXiv:2409.01652, 2024

    2024 arXiv

  7. [15]

    P. Liu, Y . Orru, C. Paxton, N. M. M. Shafiullah, and L. Pinto. Ok-robot: What really matters in integrating open-knowledge models for robotics.arXiv preprint arXiv:2401.12202, 2024

    2024 arXiv

  8. [16]

    L. X. Shi, B. Ichter, M. Equi, L. Ke, K. Pertsch, Q. Vuong, J. Tanner, A. Walling, H. Wang, N. Fusai, et al. Hi robot: Open-ended instruction following with hierarchical vision-language- action models.arXiv preprint arXiv:2502.19417, 2025

    2025 arXiv

  9. [17]

    Y . Dai, J. Lee, N. Fazeli, and J. Chai. Racer: Rich language-guided failure recovery policies for imitation learning.arXiv preprint arXiv:2409.14674, 2024

    2024 arXiv

  10. [18]

    Huang, Y

    S. Huang, Y . Liao, S. Feng, S. Jiang, S. Liu, H. Li, M. Yao, and G. Ren. Adversarial data col- lection: Human-collaborative perturbations for efficient and robust robotic imitation learning. arXiv preprint arXiv:2503.11646, 2025

    2025 arXiv

  11. [19]

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakr- ishnan, K. Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691, 2022

    2022 arXiv

  12. [20]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, X. Chen, K. Choromanski, T. Ding, D. Driess, A. Dubey, C. Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control.arXiv preprint arXiv:2307.15818, 2023

    2023 arXiv

  13. [21]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware.arXiv preprint arXiv:2304.13705, 2023

    2023 arXiv

  14. [22]

    X. Li, M. Liu, H. Zhang, C. Yu, J. Xu, H. Wu, C. Cheang, Y . Jing, W. Zhang, H. Liu, et al. Vision-language foundation models as effective robot imitators.arXiv preprint arXiv:2311.01378, 2023

    2023 arXiv

  15. [23]

    Ghosh, H

    Octo Model Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, C. Xu, J. Luo, T. Kreiman, Y . Tan, L. Y . Chen, P. Sanketi, Q. Vuong, T. Xiao, D. Sadigh, C. Finn, and S. Levine. Octo: An open-source generalist robot policy. InProceedings of Robotics: ...

    2024

  16. [24]

    Haldar, Z

    S. Haldar, Z. Peng, and L. Pinto. Baku: An efficient transformer for multi-task policy learning. arXiv preprint arXiv:2406.07539, 2024

    2024 arXiv

  17. [25]

    Nasiriany, S

    S. Nasiriany, S. Kirmani, T. Ding, L. Smith, Y . Zhu, D. Driess, D. Sadigh, and T. Xiao. Rt-affordance: Affordances are versatile intermediate representations for robot manipulation. arXiv preprint arXiv:2411.02704, 2024

    2024 arXiv

  18. [26]

    AgiBot-World-Contributors, Q. Bu, J. Cai, L. Chen, X. Cui, Y . Ding, S. Feng, S. Gao, X. He, X. Hu, X. Huang, S. Jiang, Y . Jiang, C. Jing, H. Li, J. Li, C. Liu, Y . Liu, Y . Lu, J. Luo, P. Luo, Y . Mu, Y . Niu, Y . Pan, J. Pang, Y . Qiao, G. Ren, C. Ruan, J. Shan, Y . Shen, C...

    2025 arXiv

  19. [27]

    Bjorck, F

    J. Bjorck, F. Casta ˜neda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    2025 arXiv

  20. [28]

    Y . Hu, F. Lin, T. Zhang, L. Yi, and Y . Gao. Look before you leap: Unveiling the power of gpt-4v in robotic vision-language planning.arXiv preprint arXiv:2311.17842, 2023

    2023 arXiv

  21. [29]

    A. Xiao, N. Janaka, T. Hu, A. Gupta, K. Li, C. Yu, and D. Hsu. Robi butler: Remote multimodal interactions with household robot assistant.arXiv preprint arXiv:2409.20548, 2024

    2024 arXiv

  22. [30]

    L. X. Shi, Z. Hu, T. Z. Zhao, A. Sharma, K. Pertsch, J. Luo, S. Levine, and C. Finn. Yell at your robot: Improving on-the-fly from language corrections.arXiv preprint arXiv: 2403.12910, 2024

    2024 arXiv

  23. [31]

    Belkhale, T

    S. Belkhale, T. Ding, T. Xiao, P. Sermanet, Q. Vuong, J. Tompson, Y . Chebotar, D. Dwibedi, and D. Sadigh. Rt-h: Action hierarchies using language.arXiv preprint arXiv:2403.01823, 2024

    2024 arXiv

  24. [32]

    B. Xiao, H. Wu, W. Xu, X. Dai, H. Hu, Y . Lu, M. Zeng, C. Liu, and L. Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks (2023).URL https://arxiv. org/abs/2311.06242, 2023

    2023 arXiv

  25. [33]

    M. Ding, B. Xiao, N. Codella, P. Luo, J. Wang, and L. Yuan. Davit: Dual attention vision transformers. InEuropean conference on computer vision, pages 74–92. Springer, 2022

    2022

  26. [34]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polo- sukhin. Attention is all you need.Advances in neural information processing systems, 2017

    2017

  27. [35]

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone. Libero: Benchmarking knowl- edge transfer for lifelong robot learning.arXiv preprint arXiv:2306.03310, 2023

    2023 arXiv

  28. [36]

    A. Z. Ren. open-pi-zero.https://github.com/allenzren/open-pi-zero, 2024

    2024

  29. [37]

    Lewis, Y

    M. Lewis, Y . Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V . Stoyanov, and L. Zettlemoyer. Bart: Denoising sequence-to-sequence pre-training for natural language gen- eration, translation, and comprehension.arXiv preprint arXiv:1910.13461, 2019

    1910 arXiv

  30. [38]

    pick up plate and place on plate

    OpenAI. Hello gpt-4o.https://openai.com/index/hello-gpt-4o/, 2025. 11 Appendix A Overview This appendix provides comprehensive technical details of SwitchVLA, organized into several key aspects. Section B presents implementation specifics, including: (1) our dynamic condition-...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.