Paper 1
■書誌情報
Kohei Sendai, Tatsuya Matsushima, Yusuke Iwasawa: ANVIL: System Integration for Autonomous VLA Policy Refinement — A Modular Real-Robot Loop and Empirical Lessons, The 2027 IEEE/SICE International Symposium on System Integration (SII 2027), January 2027
■概要
Vision-language-action (VLA) policies are normally improved by collecting more human teleoperation. We present an auditable, modular system that closes the real-robot collect–judge–train–deploy loop with no human in the path, and use three instantiations to identify training-signal transformation and evaluation calibration as the dominant practical failure modes. Handed a π0.5 base finetuned for 5k steps, five autonomous rounds on a physical SO-101 arm (260 rollouts, 6.9 h of arm time, five remote finetunes) returned a checkpoint that matches the base where it was already strong and is markedly better where it rarely succeeded (1/10 → 7/10; pooled 42.0 % → 60.0 %), while a demonstrations-only control trained for twice as many steps does not reproduce the gain (36.0 %). The system’s learning-critical components each sit behind one narrow, replaceable interface — policy inference behind a socket, a success/fail signal from a VLM judge, and an embodiment of dora dataflow nodes — and the interfaces were exercised in practice: when reward-weighted behaviour cloning failed, an offline audit identified an uninformative chunk-level training signal, and switching to success-only finetuning required changing one configuration key. The evidence is one evaluation session, and the loop’s own protocol is not calibrated against manual evaluation (the unchanged base reads 22.0 % in-loop but 42.0 % manually); both limits are analysed in the paper.
Paper 2
■書誌情報
Kohei Sendai, Tatsuya Matsushima, Yusuke Iwasawa: MINERVA: How Small Can a Manipulation Policy Be and Still Solve LIBERO?, The 2027 IEEE/SICE International Symposium on System Integration (SII 2027), January 2027
■概要
Vision-language-action (VLA) models with billions of parameters now dominate the LIBERO manipulation benchmark, yet how much capacity the benchmark actually demands has not been measured. From a deployment standpoint this is the quantity that matters: once a robot’s task set is fixed, the smallest policy that still solves it governs deployment cost — for example, whether the policy fits on an edge device. We present MINERVA (MINimal Efficient Robotic Vision-Action policy), a family of deliberately minimal visuomotor policies and scale it down until it breaks. A 0.54M-parameter policy reaches 95.1% average success over 2,000 rollouts on the four standard LIBERO suites, 2.4 points below the reported LeRobot π0.5 result from a model 7,700× larger; performance saturates near one million parameters and collapses below a quarter million. Across a broad sweep of architectural, training, and inference choices, only action-chunk length and vision allocation consistently exceed a ±1-point training-seed band. Flow matching shows no detectable advantage over direct L1 regression across three seeds, while regression is up to 3.8× faster on GPU. A permutation probe makes the role of task conditioning causal: rewriting the task-ID mapping alone collapses success to near chance, showing that instruction conditioning on standard LIBERO primarily selects among memorized tasks. The same recipe reaches 94.6% over 89 LIBERO-90 tasks in a single-seed extension, while LIBERO-Plus evaluation drops success to 46–56% under perturbations, with near-zero photometric robustness across all tested scales. The 0.54M policy replans at every control step in 5–9 ms per chunk on a laptop CPU, 113× faster than SmolVLA and 1,400× faster than π₀.₅, without a GPU. These results establish a first empirical estimate of LIBERO’s task-specific capacity floor and motivate capacity-aware design and distillation for deployment-efficient robot policies.
Paper 3
■書誌情報
Makoto Sato, Tatsuya Matsushima, Yutaka Matsuo, Yusuke Iwasawa: DREAM: Deployment-Time Demonstration Generation via Real-to-Sim for Scalable Policy Adaptation, The 2027 IEEE/SICE International Symposium on System Integration (SII 2027), January 2027
■概要
Vision-language-action (VLA) models have made strong progress in language-conditioned robot manipulation, but improving their performance in a new workspace still often requires action-labeled data from that environment. Collecting such data by human teleoperation is costly, especially when each workspace, object arrangement, or task may require new demonstrations. We present DREAM, a framework that generates fine-tuning data for a pretrained VLA from a captured workspace and a language instruction, without requiring a task-specific human demonstration. DREAM reconstructs the workspace, automatically translates the instruction into symbolic task goals and success criteria using a large language model, and uses task-and-motion planning to generate feasible robot trajectories. The planned trajectories are augmented across randomized object configurations, verified by the generated success criteria, and rendered into image-action examples for VLA fine-tuning. Through real-robot experiments on language-conditioned manipulation tasks, we study whether DREAM can serve as a scalable data-collection system for the deployment workspace by examining whether fine-tuning on its automatically generated data improves success over direct deployment and how its data-collection cost compares with human teleoperation when adapting a VLA to a new workspace.
Paper 4
■書誌情報
Sodtavilan Odonchimed, Tatsuya Matsushima, Simon Holk, Yusuke Iwasawa, Yutaka Matsuo: Retrieve-Augmented Generation for Speeding up Diffusion Policy without Additional Training, The 2027 IEEE/SICE International Symposium on System Integration (SII 2027), January 2027
■概要
Diffusion Policies (DPs) have attracted attention for their ability to achieve significant improvements in accuracy across various imitation learning tasks. However, DPs rely on Diffusion Models, which require multiple noise-removal steps to generate a single action, leading to long generation times. To address this problem, knowledge distillation-based methods, such as Consistency Policy (CP), have been proposed. However, these methods require significant training time, especially for difficult tasks. In this study, we propose RAGDP (Retrieve-Augmented Generation for Diffusion Policies) as a novel framework that eliminates the need for additional training using a knowledge base to expedite the inference of pre-trained DPs. In concrete, RAGDP encodes observation-action pairs using the DP encoder to construct a vector database of expert demonstrations, enabling online retrieval. During inference, the current observation is embedded, and the most similar expert action is extracted. This extracted action is combined with an intermediate noise-removal step to reduce the number of steps required compared with the original diffusion step. We show that by using RAGDP with the base model and existing acceleration methods, we improve the accuracy-speed trade-off without additional training. Even when accelerating the models 20 times, RAGDP maintains an advantage in accuracy, with a 7% increase over distillation models such as CP.
Paper 5
■書誌情報
Ryosuke Takanami, Tatsuya Matsushima, Yutaka Matsuo, Yusuke Iwasawa: Roll Out, Intervene, and Self-Imitate: Human-in-the-Loop Rollout Finetuning for Embodiment Adaptation of Pretrained Visual Navigation Models, The 2027 IEEE/SICE International Symposium on System Integration (SII 2027), January 2027
■概要
General Navigation Models (GNMs), pretrained on large-scale multi-robot datasets, can navigate autonomously even in environments they have never seen. Deploying a GNM on a new robot whose embodiment lies outside the pretraining distribution still requires adaptation. The common approach, full-teleoperation supervised fine-tuning (FT-SFT), adapts the policy using only human-operated trajectories; this is expensive and can overwrite the pretrained behavior underlying GNMs’ generalization. We propose Human-in-the-Loop Rollout Finetuning (HIL-RFT) as an alternative. In HIL-RFT, the pretrained policy rolls out autonomously, a human intervenes only when necessary, and the entire rollout supervises fine-tuning. Autonomous segments provide self-imitation signals that anchor the model to its pretrained behavior, while intervention segments correct embodiment-specific failures. FT-SFT and HIL-RFT thus share the base policy, objective, and hyperparameters, differing mainly in how the adaptation data are collected and composed. Using NoMaD as a representative GNM in Gazebo, we compare them under an approximately matched training-frame budget, with zero-shot NoMaD as a reference. HIL-RFT reduces the human-generated share of training frames from 100% to 65.8%, while achieving a higher in-domain success rate than FT-SFT (75.0% vs. 37.5%). On a held-out environment, FT-SFT does not improve over zero-shot NoMaD (50% to 50%), whereas HIL-RFT reaches 75%. These results suggest HIL-RFT as a low-human-effort adaptation strategy where preserving the policy’s own rollout behavior matters as much as collecting corrective human actions.
Paper 6
■書誌情報
Ryosuke Takanami, Tatsuya Matsushima, Yaonan Zhu, Masato Kobayashi, Yutaka Matsuo, Yusuke Iwasawa: Prompt-Correctable Open-Vocabulary 3D Semantic Maps for Shared-Control Teleoperation, The 2027 IEEE/SICE International Symposium on System Integration (SII 2027), January 2027
■概要
Mobile manipulators are difficult to teleoperate because an operator must coordinate a mobile base and a multi-degree-of-freedom arm from incomplete remote observations. We present a shared-control interface that back-projects open-vocabulary instance masks into an operator-visible 3D semantic map. The operator selects a mapped object either spatially with a virtual-reality controller or by its label, after which the robot autonomously coordinates its base and arm to reach a fixed pre-grasp pose. When the detector’s generic vocabulary does not distinguish deployment-specific objects, the operator can supply short descriptions that prompt the detector vocabulary and an external vision-language re-classifier at run time, without gradient updates. We assess the two system functions separately. In a counterbalanced pilot study with six operators, reach assistance reduced the mean completion time for an easy pick-up target by 29.2 s and for a two-object tidying task by 63.5 s, with large paired effect sizes but without confirmatory statistical evidence. In a descriptive evaluation on 12 object instances, the best prompt configuration raised map-label accuracy from 33.3 to 91.7 percent and the success rate of the resulting label-guided motions from 25.0 to 91.7 percent. These pilot results support the feasibility of combining operator-correctable semantic mapping with shared-control teleoperation while identifying grasp-pose estimation, prompt-authoring cost, and larger multi-environment studies as remaining work.
Paper 7
■書誌情報
Shuai Mao, Zhongchao Zhou, Konstantin Dzavaryan, Yixing Lei, Wenze Zhang, Mauricio Villarroel, Liang He: Effects of Pneumatic Artificial Muscle Layouts on Workspace and Stiffness of Modular Soft Continuum Robots, The 2027 IEEE/SICE International Symposium on System Integration (SII 2027), January 2027
■概要
Soft continuum robots can continuously deform and adapt to their surroundings, attracting broad research interest. Although McKibben pneumatic artificial muscles (PAMs) are widely used as actuators in modular soft continuum robots, the effects of PAM layouts on key characteristics of the overall robot, such as workspace and stiffness, have not been systematically investigated. In this work, we develop a modular soft continuum robot together with a reduced-order kinematic model to interpret the effects of muscle layout on robot deformation. The effects of layout variations on workspace and stiffness are then systematically investigated experimentally, considering the muscle-to-spine distance d, normal angle θ, tangential angle α, and muscle arrangement patterns (II-shape, Λ-shape, and X-shape). The results show that increasing α from 0° to 24.3° increases the maximum planar radial reach by 46.0%, while increasing d from 52 mm to 92 mm reduces it by 35.5%, and increasing θ from 0° to 11.7° reduces it by 23.5%. Changing the muscle arrangement from II-shape to Λ-shape increases the maximum planar radial reach by 59.4%. Under a 200 g load, increasing d and α reduces the measured stiffness by 40% and 77.1%, respectively, while increasing θ increases it by 56.2%. The Λ-shape provides the largest maximum planar radial reach, while the X-shape provides the largest maximum axial reach and stiffness. These findings provide a design basis for the system integration and task-oriented design of modular soft continuum robots, enabling PAM layouts to be selected according to application-specific workspace and stiffness requirements.