Research

Publications|Research Experience|Engineering Experience|Education|Scholarships & Fellowships|Academic Activities|CV (PDF)

Publications

1. PredVLA: Predictive Sensorimotor Modeling for Sub-Million-Parameter Robot Manipulation

   Hiroki Sawada and Shunichi Kasahara

   arXiv preprint (arXiv:2608.26673), 2026|Link

   Proposed PredVLA, a compact recurrent controller organized around predicting sensorimotor dynamics and inferring its latent state from prediction error, rather than compressing a conventional observation-to-action policy. With no robot-data pretraining and about 0.68M trainable parameters, it reaches 86.94% mean success on the three short-horizon LIBERO suites (75.35% across all four), against 19.73% for BC-Transformer and 10.26% for BC-LSTM under a matched parameter budget and an identical frozen front end.

   PredVLA · Predictive Coding · Robot Manipulation · LIBERO

2. Beyond Single-Axis Testing: Paired Evaluation of Compound Robustness in Vision-Language-Action Policies

   Hiroki Sawada and Shunichi Kasahara

   arXiv preprint (arXiv:2609.15940), 2026|Link

   Introduced LIBERO-CTRL, a six-axis benchmark that pairs each initial state across single-axis conditions and a matched simultaneous condition. The design exposes two opposing outcome changes that aggregate success rates cannot distinguish, emergent failures and compensated successes; because they cancel, as many as 29.0% of matched initial states change outcome even when the aggregate appears consistent with single-axis measurements, reaching 34.5% in the most affected condition across six policies and three severity levels.

   LIBERO-CTRL · VLA · Robustness · Benchmark

3. Predictive-Coding-Based Autonomous Regulation of Internally Generated and Externally Coupled Processing in Human-Robot Interaction

   Henrique Oyama, Hiroki Sawada, and Jun Tani

   arXiv preprint (arXiv:2609.06888), 2026|Link

   Proposed an online mechanism that switches the meta-prior of a PV-RNN according to reconstruction error accumulated over recent interaction history, letting the robot itself regulate how strongly inference relies on internally generated predictions versus external sensory evidence. Across three physical human-robot interaction tasks, reconstruction-history-driven regime selection reduced prospective prediction error and robot-side physical interaction conflict.

   PV-RNN · Meta-Prior · Predictive Coding · HRI

4. Free-Energy-Gated Plasticity for Real-Time Online Motor Learning in Physical Human-Robot Interaction

   Hiroki Sawada and Jun Tani

   arXiv preprint (arXiv:2608.23000), 2026|Link

   Extended PV-RNN to continuously adapt its synaptic weights and proposed Free-Energy-Gated Plasticity (FEGP), which regulates the effective learning rate according to variational free energy. A randomly initialized network acquired three cyclic motor patterns during real-time physical human-robot interaction without pretraining, replay, or task-boundary signals, showing that the timing of plasticity relative to model-environment mismatch, rather than its average magnitude, preserves previously acquired behaviors.

   FEGP · PV-RNN · Online Learning · Torobo

5. Human-Robot Kinaesthetic Interaction Based on The Free Energy Principle

   Hiroki Sawada, Wataru Ohata, and Jun Tani

   IEEE Transactions on Systems, Man and Cybernetics: Systems, 2024|Link

   Modeled physical human-robot kinaesthetic interaction using a predictive-coding-inspired variational RNN based on the Free Energy Principle. Quantitatively analyzed how the hyperparameter balancing prediction accuracy and model complexity affects human-robot force interaction.

   FEP · PV-RNN · HRI · Torobo

6. Incremental Learning in Physical Human-Robot Interaction Using Predictive-Coding inspired Variational RNNs

   Hiroki Sawada and Jun Tani

   IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2024, Late Breaking Poster|PDF

   Extended PV-RNN to incremental learning in physical HRI. Principal component analysis on the internal state space clarified how incremental learning reshapes the distribution of learned motion patterns.

   Incremental Learning · Predictive Coding · PV-RNN · Torobo

7. CERNet: Class-Embedding Predictive-Coding RNN for Unified Robot Motion, Recognition, and Confidence Estimation

   Hiroki Sawada, Alexandre Pitti, and Mathias Quoy

   2026 IEEE International Conference on Robotics and Automation (ICRA), pp. 21744-21751|Link

   Proposed CERNet, a hierarchical predictive-coding RNN with class embedding, enabling unified real-time motion generation, recognition, and meta-inference of confidence. Demonstrated the model using the Reachy humanoid robot in both simulation and real-world settings.

   CERNet · ICRA 2026 · Predictive Coding · Reachy

8. Anomalous Magnetic Moment Direction under Magnetic Anisotropy Originated from Crystalline Electric Field in van der Waals Compounds CeTe3 and CeTe2Se

   Ueta Daichi, Riki Kobayashi, Hiroki Sawada, et al.

   Journal of the Physical Society of Japan 91(9):094706, 2022|Link

   Investigated magnetic anisotropy and anomalous magnetic moment direction in van der Waals compounds CeTe3 and CeTe2Se.

   Condensed Matter · Magnetism

9. In-gap-states of a mixed anion layered compound, polycrystalline LaCu1-deltaS0.5Se0.5O(delta~0.01) as a degenerate semiconductor

   Azuma Nobuhiko, Hiroki Sawada, Itoh Houji, Sakagami Ryosuke, Matoba Masanori, Usui Hidenori, and Kamihara Yoichi

   Journal of the Society of Materials Science, Japan 58, 64-68, 2021|PDF

   Studied in-gap states in the mixed-anion layered compound LaCu1-deltaS0.5Se0.5O as a degenerate semiconductor.

   Materials Science · Thermoelectric

10. Thermal transport properties of a mixed anion layered compound, polycrystalline LaCu1-δS0.5Se0.5O (δ = 0.01)

    N. Azuma, H. Sawada, H. Ito, R. Sakagami, Y. Tanaka, T. Fujioka, M. Matoba, and Y. Kamihara

    Korean Journal of Materials Research 34, 464-474, 2024|Link

    Investigated thermal transport properties of the mixed-anion layered compound LaCu1-δS0.5Se0.5O.

    Materials Science · Thermal Transport

Research Experience

Engineering Experience

Education

Scholarships & Fellowships

Academic Activities