Multimodal Neurotechnology Group Institute of Cognitive Neuroscience and Psychology HUN-REN Research Centre for Natural Sciences Hungarian Research Network
Highway lane-change maneuvers pose a safety challenge for SAE Level 3 autonomous driving because conventional threshold-based decisions and isolated surrogate safety measures may inadequately capture the coupling between longitudinal braking demand and evolving inter-vehicle gaps. This study develops an interpretable, risk-aware framework for continuous lane-change assessment during maneuver initiation and execution. An adaptive neuro-fuzzy inference system (ANFIS) combines normalized longitudinal braking demand and minimum inter-vehicle gap to generate a continuous risk index ranging from zero to one. The index is integrated into a two-phase decision mechanism incorporating multi-vehicle threat assessment, region-dependent maneuver cancellation, quintic Bézier trajectory replanning, and driver takeover supervision. Performance was evaluated through high-fidelity SCANeR/ReScan and MATLAB/Simulink co-simulations against static Mamdani fuzzy, time-to-collision (TTC), deceleration-demand, and time-headway baselines. Robustness was assessed using 1,000 Monte Carlo simulations with variations in vehicle speeds, sensor noise, road friction, and cut-in timing. Compared with the TTC baseline, the proposed framework increased warning lead time from 0.86 to 1.84 s and minimum two-dimensional clearance from 2.08 to 4.82 m, while reducing maximum deceleration demand from 4.92 to 2.34 m / s 2 . Mean computational execution time was 1.42 ms, below the 10-ms simulation control-cycle budget. Across the 1,000 randomized simulations, no collisions were observed with the proposed framework; near-miss and false-alarm rates were 0.30% and 1.20%, respectively. The simulation findings support the potential of continuous neuro-fuzzy risk assessment to provide earlier hazard detection and larger safety margins while retaining rule-based interpretability and low computational cost. Hardware-in-the-loop and real-vehicle testing remain necessary to establish performance beyond the evaluated simulation conditions.
Bimanual robotic manipulation remains a significant challenge in embodied AI, as it requires synchronized motion planning, inter-arm collision avoidance, and dynamic role allocation over long task horizons. Although existing diffusion-based policies and affordance learning methods have demonstrated considerable capability, they largely treat the environment as static and fail to model how interaction relationships and coordination patterns evolve over time. This paper introduces the Predictive Affordance and Coordination Evolution Diffusion Policy (PACE-DP), a unified framework that explicitly forecasts future scene dynamics and models the temporal progression of manipulation affordances and bimanual coordination. PACE-DP integrates three predictive components: (1) a Predictive Scene Module that forecasts latent future object-centric representations; (2) an Affordance and Coordination Evolution Module that constructs temporally evolving interaction graphs and employs a Graph Transformer to predict future relation types, grasp success probabilities, coordination modes, and spatial alignment targets; and (3) a Conditional Diffusion Policy that generates smooth, coordinated action sequences conditioned on fused current and predicted features. We evaluate PACE-DP on the RoboTwin 2.0 platform across 10 diverse bimanual tasks, including handover, placement, stacking, and switch operation. Experimental results show that PACE-DP performs favorably compared with baseline methods, achieving higher success rates on coordination-intensive tasks with only marginal computational overhead. These findings suggest that explicitly predicting future affordance and coordination evolution provides an effective inductive bias for bimanual manipulation, enabling more reliable coordination without substantially increasing computational cost. PACE-DP thus offers a promising step toward predictive, temporally aware policies for long-horizon bimanual tasks.
Deep Reinforcement Learning (DRL), which integrates reinforcement learning with deep neural networks (DNNs), has been extensively researched across diverse domains. Robotics, in particular, has seen significant advancements, as DRL enables agents to extract meaningful features from high-dimensional observations and make precise decisions within complex environments. Furthermore, in real-world control tasks, achieving high performance within a limited number of interactions is required, particularly in complex and unknown environments. In addition, since real-world trials are constrained by cost and safety considerations, designing DRL algorithms that enable high sample efficiency and stable decision-making is of critical importance. Spiking neural networks (SNNs), inspired by biological neural systems, have recently attracted increasing attention due to the binary and event-driven nature of their information processing. Motivated by these properties, there has been growing interest in integrating SNNs into DRL. Nevertheless, SNN-based DRL faces challenges in learning performance. To address this issue, this study introduces embeddings learned from environmental dynamics into SNN-based DRL. Experimental evaluations conducted on four continuous control tasks from OpenAI Gym demonstrate that the proposed method outperforms conventional SNN-based DRL methods, both in terms of maximum average rewards and accelerated reward acquisition during the early stages of learning. These results establish that leveraging transition-aware embeddings effectively enhances both performance and sample efficiency in SNN-based DRL.
Automated posture quality assessment has become increasingly important in health and wellness technology. However, existing methods predominantly provide point estimates without quantifying prediction confidence, posing a critical limitation in safety-sensitive applications where incorrect assessments may lead to improper guidance. This paper addresses this gap by introducing a reliability-oriented evaluation framework based on Bayesian Liquid Neural Networks (LNNBI). Unlike conventional approaches, our method simultaneously delivers a quality score and a principled uncertainty measure derived from Bayesian variational inference. More importantly, we incorporate a rejection mechanism that enables the model to abstain from making predictions when uncertainty exceeds a predefined threshold, establishing a built-in safeguard for practical deployment. To support this framework, we construct a longitudinal dataset comprising 50 participants over 200 sessions, decomposing four fundamental human postures into seven biomechanical indicators. Experimental results demonstrate that LNNBI achieves 0.8715 accuracy and 0.9105 precision, outperforming LSTM, MC-GRU, TSGCN, and YoNet baselines. Uncertainty analysis reveals a strong correlation between prediction error and uncertainty (Pearson’s r = 0.73), confirming that the model reliably identifies its own low-confidence predictions. With a 15% rejection rate, the effective accuracy reaches 0.923. Correlation analysis further identifies Pelvic Position and Neutrality ( r = 0.682) and Spinal Alignment ( r = 0.595) as the most influential biomechanical indicators. This work offers a reliability-first perspective on automated posture assessment, with promising potential for deployment in personalized health coaching and remote guidance, particularly on GPU-accelerated systems.
Vision-Language-Action (VLA) models offer a unified interface for language-conditioned manipulation. This paper reports a negative result: a low-cost post-training recipe for SmolVLA with LoRA, a temporal smoothness regularizer, and chunk-aware auxiliary supervision from action-variation pseudo labels does not improve manipulation success under a 30,000-step budget (approximately 1/67 of the original 2,000,000-step SmolVLA recipe). Because the auxiliary classifier is used only during training and the runtime chunk length remains fixed at k = 8, we denote the configuration SmolVLA+LoRA+S&C (smoothness and chunk-aware auxiliary losses) rather than “adaptive.” We evaluate four LIBERO suites with a unified protocol of 500 episodes per cell (50 episodes × 10 task variants). Over three training seeds, the matched fixed-chunk k = 8 baseline reaches a four-suite macro success rate of 10.9%±1.1, whereas SmolVLA+LoRA+S&C reaches 3.5%±1.4; the pre-specified primary comparison shows significant deficits on LIBERO-Spatial (−16.4 percentage points, exact sign-flip permutation p = 0.004) and LIBERO-Object (−17.5 pp, p = 0.002), with matched-pairs rank-biserial effect sizes of −0.96 and −1.0. Matched component ablations isolate the smoothness loss as the harmful term (smoothness-only collapses performance; chunk-classifier-only approximately matches the baseline), and executed-action logs show the regularizer triples trajectory jerk (1.69 vs. 0.59 RMS)—the opposite of its intent. A training-free inference-time temporal-ensembling baseline, by contrast, more than doubles success on the two suites where the baseline is competent (Spatial 36.2 vs. 19.0%; Object 56.2 vs. 24.0%) at a 2.3 × step-rate cost, and a 10–100K budget sweep shows the S&C deficit never closes. We conclude that, for the tested configuration, budgets, and benchmark, the added action-side losses are not competitive with a well-tuned fixed chunk, and inference-time ensembling is the stronger use of the same compute.
