Defense Day

Time and Place

The defense session will take place on September 25, 2026, Friday at 10:00 AM. 

You can join the Microsoft Teams meeting room using the information below.

Microsoft Teams meeting

Join: https://teams.microsoft.com/meet/354902393026653?p=hOwVE4PVm69TC99mIr
Meeting ID: 354 902 393 026 653
Passcode: gz6G5kN7

Hosted by Yinglei Song (yinglei.song@tum.de)

Schedule

10:00 - 10:20 Paul Fessele (BA Thesis)

Title: Bimanual Manipulation on a Humanoid Robot: Quantifying Speedup and Coordination Constraints

Advisor: Zitao Zhang

Keywords:

Abstract: This thesis studies the speedup of parallel bimanual cube stacking on a humanoid robot. A modular perception–planning–execution pipeline is used to compare single-arm, sequential bimanual, and parallel bimanual execution on identical scenes. With ground-truth perception, parallel execution achieves a median speedup of up to 1.56 when the arms build separate towers, but only 1.22 for a shared tower. The reduced speedup is mainly caused by dependent placements and competition for the shared waist. Vision-language-model perception only slightly reduces task success, but increases total execution time due to blocking perception calls. The results show that bimanual manipulation is most effective when the task can be divided into largely independent subtasks.

10:20 - 10:40 Huang, Yiyang (BA Thesis)

Title: Vision-Language Guidance for Target-Oriented Robotic Grasping in Cluttered Scenes

Advisor: Tianyao Tong

Keywords: Vision-Language Model; Robotic Grasp Detection

Abstract: Target-oriented robotic grasping in cluttered scenes requires a robot to identify the intended object from a language instruction and select a feasible grasp in a complex visual scene. This thesis studies how a vision-language model (VLM) can guide this process in simulation. The proposed pipeline uses the VLM to provide structured target information, which is combined with GroundingDINO for open-vocabulary localization and GraspNet for 6-DoF grasp generation. The selected grasp is then executed by a simulated Franka Emika Panda robot arm with execution monitoring and re-planning. The aim of this thesis is to examine whether this guidance can better connect natural-language target instructions, cluttered visual observations, and executable grasp actions.

10:40 - 11:00 Wang, Simou (BA Thesis)

Title: Meta-Agent Platform for Multi-Agent Workflow Automation

Advisor: Tianyao Tong

Keywords: Multi-Agent Workflow; Self Evolving

Abstract: This thesis designs and implements a Meta-Agent Platform for business process automation. The main contribution of this thesis is the system design and implementation of the platform. The platform consists of five core modules: Requirement Clarification & Agent/Workflow Generation, Multi-Agent Workflow Execution, External Tool Integration, Execution Monitoring & Logging, and Process Mining. These modules are connected through one workflow lifecycle and form the complete platform architecture. To evaluate the prototype, this thesis uses 24 test cases and performs 120 Workflow Runs. The tests cover Single-Agent, Multi-Agent, external-tool, and conditional-execution scenarios. The results show that the platform can generate and execute Workflows with different levels of complexity, coordinate multiple Agents and external tools, and use real execution data for monitoring and process analysis.

11:00 - 11:25 Juan Victor Sandalinas (MA Thesis)

Title: Generative-Based Occlusion-Aware Trajectory Prediction with Multi-Sensor Fusion

Advisor: Yingjie Xu

Keywords: Autonomous driving Safety, Occlusion Prediction, Flow Matching

Abstract: Motion forecasting is an essential field for safe autonomous driving, yet current models assume perfect knowledge of agents’ past trajectories and knowledge of HD maps, which in reality might not be available. Specifically, the latter assumption does not hold in the case of occluded scenarios, where an agent’s full past track is not available. This thesis presents an occlusion-based generative model architecture that leverages Radar and Camera data for correct identification and velocity estimation by the use of BEV features and a generative decoder. We use the RCBEVDet encoder for Radar–Camera BEV features, a Scene Context Encoder for fusion within the encoder, and two decoder options: a Flow-Matching-based iterative cross-attention and deformable cross-attention decoder that further leverages BEV feature map exploration, and a diffusion-based decoder. The proposed methods are evaluated in the nuScenes dataset benchmark by generating synthetic occlusions derived from the camera visibility tokens. Experimental results showed a decrease in performance in cross-attention methods when increasing exploration up to 87% of the BEV map, with an error increase of 7.8% in minADE and 8.5% in minFDE, and how the advantages of using deformable cross-attention did not show relevant degradation in performance, whilst achieving a 1.65× speedup in FPS and a 3.4× reduction in VRAM usage against cross attenion. We further compare our methods 2 generative methods against the nuScenes challenge occluded benchmark, where we showed that the diffusion framework without cross-attending diffusion layers did not show any improvement and was not able to capture the BEV feature space. In contrast, the Flow Matching decoder offered a 3.2× speedup, with a 200 Mb increase, beat the baseline TrajFlow HD-map-free by 4.7% in minADE (up to 6.2% with cross-attention), reduced occluded velocity error (mAVE) by 7.8%, and improved minADE by 9.1%.

11:25 - 11:45 Jakub Jura (BA Thesis)

Title: Using Differentiable Simulation for Quadrupped Locomotion

Advisor: Yulong Xiao

Keywords: Quadruped Locomotion

Abstract: This thesis investigates differentiable simulation for quadruped locomotion on challenging terrain. The work extends an existing differentiable locomotion framework for the Unitree Go2 by introducing full-pipeline vectorization, a fused CUDA kernel with an analytical adjoint, differentiable terrain height sampling, privileged height observations, and a simulated depth camera. The resulting implementation enables large-scale training, providing up to $12$x speed-up over the inherited code. Further experiments showed that neither terrain gradients nor terrain observations yielded a reliable improvement in hard-terrain locomotion. Across the training runs, the terrain curriculum consistently collapsed to its easiest level, limiting the policies' ability to learn terrain-dependent behavior, mainly due to instability under omnidirectional commands and the absence of a direct penalty for falling. Several interventions were evaluated, but none prevented the collapse.

11:45 - 12:05 Guojun Rao (BA Thesis)

Title: Real-Time Embedded Fisheye Distortion Correction on FPGA for Robot Vision Systems

Advisor: Qian Huang

Keywords: Fisheye cameras, Hardware Acceleration, Rendering, FPGA

Abstract: Fisheye cameras provide a wide field of view but use a nonlinear projection that is incompatible with many pinhole-based vision algorithms. Existing FPGA implementations often rely on pixel-wise lookup tables, resulting in considerable storage and memory-access overhead. This thesis develops a computation-based forward-mapping architecture that replaces the complete mapping table with a fixed-point and deeply pipelined datapath. For each incoming fisheye pixel, the system computes its position in one of two virtual pinhole views. A seventh-order polynomial camera model is inverted using two fixed Newton-Raphson iterations, while scale-equivalent ray reconstruction and fixed view orientations simplify the geometric computation. Subsequent stream-straightening and 3 × 3 densification stages convert the irregular and sparse mapping results into raster-ordered images. RTL simulations and Python reference models confirm the numerical and functional correctness of the design. The mapping core has a deterministic latency of 270 clock cycles and accepts one sample per cycle after pipeline filling. Consecutive 1024 × 960 frames were transformed into two 500 × 384 views per frame. The results demonstrate the feasibility of replacing complete pixel-wise mapping storage with real-time arithmetic computation.

12:05 - 12:30 Moritz Heber (MA Thesis)

Title: Hydrodynamic Modeling and Trajectory Optimization for Propulsion Efficiency in Amphibious Quadruped Robots

Advisor: Qian Huang

Keywords: Amphibious Quadruped Robot, Trajectory Optimization, Hydrodynamic Modeling, Drag-Based Paddling, Optimal Control

Abstract: Amphibious quadruped robots can swim by paddling with their legs. As paddling is dragbased, its efficiency depends heavily on the stroke. Strokes inspired by biology are not designed for the specific robot’s leg geometry and actuator limits. This thesis uses trajectory optimization over a reduced-order hydrodynamic model to obtain swimming gaits. This model applies buoyancy, drag, and added mass to each link of the partially submerged robot, while remaining differentiable as links cross the water surface. Coefficients of the model are calibrated against a smoothed particle hydrodynamics simulation. A periodic optimal control problem then minimizes the squared mechanical joint power at a prescribed minimum average forward speed. The optimal control problem is complemented by regularization terms and physical constraints of the robot platform. The calibrated model reproduces the forward displacement per cycle of the reference within 10 % on the calibration runs. Across a total of 320 optimizations between 0.10 m s−1 and 0.25 m s−1, speed and efficiency form a Pareto front that is limited by the range and speed of the joints. The optimized gait allocates approximately one third of the cycle to a fast power stroke, generating the entire net thrust with the front legs, qualitatively matching biomechanical observations. Validation on hardware remains future work.

12:35 - 13:00 Lunch Break 

13:00 - 13:25 Karim Elsharawy (MA Thesis)

Title:  AI Agent Networks and Market-Based Collaboration

Advisor: Fengjunjie Pan

Keywords: 

Abstract: Global commodity markets and multi-tier supply chains are complex adaptive systems characterized by decentralized interactions, geographic transport frictions, and vulnerability to severe geopolitical shocks. While traditional macroeconomic time-series models and numerical agent-based simulations provide aggregate numerical forecasts, they operate as black boxes, incapable of explaining the underlying behavioral rationales, strategic negotiations, and emerging operational bottlenecks of individual market participants. This thesis presents a grounded, generative multi-agent simulation framework that couples the contextual reasoning of Large Language Models (LLMs) with deterministic economic execution primitives to create an auditable, interactive supply-chain digital twin. The platform architecture integrates strict relational state transition guards, periodic sealed-bid batch call auctions, multi-tier geographic transport surcharge graphs, immutable double-entry SQLite accounting ledgers, and live RSS news ingestion pipelines. Crucially, the system addresses long-context degradation through sliding transaction observation windows and structured tool-calling protocols. Through iterative empirical stress-testing across more than 120 simulation runs, we identify and resolve four fundamental behavioral failure modes exhibited by autonomous LLM agents: (1) market trading freezes under price spikes, resolved by introducing constant elasticity of substitution (CES) demand curves with core shortage penalties; (2) bank capitalmonopolization, resolved by enforcing portfolio underwriting caps and credit rationing; (3) debt amnesia, cured by injecting high-salience default warning banners into agent observation prompts; and (4) price signal fragmentation in asynchronous continuousorder books, eliminated by transitioning to synchronized periodic batch call auctions.In an empirical crisis experiment modeling a global crude oil and refined fuel networkpowered by GLM 5.2, autonomous agents demonstrated strong economic reasonable-ness (RQ1): producers priced above extraction unit costs, refiners achieved a costpass-through elasticity of δ ≈0.82 while defending positive crack spreads ($22.00 to$41.10/bbl), and all twelve representative firms sustained a 100% solvency survival rate across 72 executed trades. To enable transparent interpretability (RQ2), we introduce apost-simulation interrogation service that rehydrates agent memory states from persis-tent relational ledgers and ChromaDB semantic vector embeddings. Natural-language debriefings with upstream producers, refiners, commercial lenders, and industrial consumers successfully extracted granular causal explanations for strategic pricing and inventory decisions. A systematic ground-truth audit (RQ3) cross-examining 120 interview claims against 261 immutable double-entry ledger records validated an overall 88.3% Factual Claim Accuracy with zero financial hallucinations. Finally, we demonstrate that conversational multi-agent simulation provides profound added value (RQ4) by uncovering hidden microeconomic mechanisms that cannot be inferred from aggregate price curves alone, establishing a generalizable foundation for next-generation supply-chain risk analysis and market mechanism design.

13:25 - 13:50 Meriem Yacoubi (MA Thesis)

Title: Memory Lifecycle in LLM Agents: Consolidation and Epistemic Pruning

Advisor: Nenad Petrovic

Keywords: 

Abstract: In this work, we aim to address the limitations of current external memory architectures for large language model agents. Existing textual memory systems can efficiently retrieve semantically relevant memories, but many of them lack the capacity to clearly define the type of relationship between two pieces of information. Although more structured approaches can represent these relationships using global graphs, hierarchical abstractions, or reflective processes, this often comes at the cost of additional complexity or computational overhead. To address this gap, we propose a novel approach, MemoryLACE (MemLACE). LACE in the name stands for Lifecycle-Aware Consolidation and Evidence Retrieval. It is a lightweight memory framework that explicitly models how textual evidence evolves through sparse merge, supersession, and contradiction relations while preserving atomic natural-language memories together with their provenance. During retrieval, these relations are used to reconstruct relation-aware evidence units containing current, historical, supporting, and conflicting evidence. To validate our approach, we evaluate MemLACE on two complementary benchmarks, BEAM and StructMemEval. MemLACE achieves the highest overall performance in the samebackbone comparisons, while reducing the end-to-end runtime on BEAM by 66.6% compared with Hindsight. These results suggest that a middle-ground framework can address some of the current limitations of such architectures.

13:50 - 14:15 Xu, Zhenzhe (MA Thesis)

Title: Design and Research of Mobile Robots for Space Environments

Advisor: Li, Yuchen

Keywords: Mobile Robots, Helium-3 Resource, Space Robotics, OS-Rover

Abstract: Future lunar helium-3 resource utilization requires mobile robotic systems capable of accessing, loosening, collecting, and transporting lunar regolith. Inspired by NASA’s ISRU Pilot Excavator (IPEx), this thesis presents the design, implementation, and ground-based validation of a low-cost rover demonstrator named OS-Rover. In this name, “OS” stands for “Open Source”, reflecting the platform’s open-source development and public dissemination. OS-Rover addresses the upstream regolith-acquisition stage of future lunar resource-utilization missions; it does not detect or extract helium-3 directly. The rover employs a four-wheel skid-steering chassis driven by four independent DJI M3508 geared motors. Each wheel incorporates six replaceable tread segments for localized maintenance and tractionsurface adaptation. Its soil-working system combines a front drill, a vertical feed, two collecting discs, and a movable arm. The drill loosens relatively compact soil before collection, whereas the discs temporarily gather the disturbed material. The movable arm and collecting discs can also provide mechanical assistance when the rover becomes locally trapped in loose terrain. An STM32F407IGT6-based controller regulates four wheel-speed loops through CAN1 and commands five soil-working actuators through CAN2. The firmware implements stopped initialization, command limits, failsafe handling, communication timeouts, and independent stop commands. Manual operation is provided through a direct SBUS radio link, while a Raspberry Pi 5 supports camera-based gesture recognition, Bluetooth command forwarding, telemetry reception, and data logging. A feedback-based recovery strategy is proposed for future implementation using wheel-speed and torque-related current feedback, but it has not yet been experimentally implemented or validated. Ground-based demonstrations on natural sand covered flat-terrain mobility with drill operation, inclinedsand traversal with mechanical assistance, and recovery from a locally trapped condition through manually coordinated actuator commands. Telemetry and video documented all three scenarios, and gesture processing operated at 41–44 frames per second. These demonstrations verify coordinated operation of the mechanical and control subsystems, but do not establish maximum mobility or excavation performance because vehicle displacement, wheel slip, interaction forces, and collected-soil mass were not independently measured. A separate handcart-based experiment evaluated a 1200 mm × 760 mm, 200W ETFE photovoltaic panel, an MPPT controller, a 24 V-class battery, and RS485/Modbus RTU data logging. The selected 237.336 s record contained 238 valid samples. Calculated battery-side charging power reached 100.23W, and the integrated charge-channel energy was 2.1267W h. Because the panel was not mechanically integrated into OS-Rover, neither MPPT efficiency nor rover endurance was determined. The experiment nevertheless provides a basis for future onboard photovoltaic integration. A paper based on this work, entitled “OS-Rover: An Open-Source Planetary Rover for Resource Prospecting with Drill-Assisted Sampling and Multimodal Teleoperation,” has been accepted for presentation at the 2026 International Conference on Space Robotics (iSpaRo 2026). The source code, mechanical design files, experimental data, and supplementary materials are publicly available at https://github.com/ OpenSpace-Lab/OS-Rover under the MIT License.

14:15 - 14:35 Helali, Anas (BA Thesis)

Title: Test-Time Training for ViT-Based Grasp Detection: Evaluation and Robustness Across the Cornell and OCID Datasets

Advisor: Hu Cao

Keywords: Robotic Grasp Detection; Test-Time Training 

Abstract: This report evaluates a Vision Transformer-based robotic grasp detection model in which the standard softmax attention is replaced by Test-Time Training (TTT) layers, and compares it against the baseline model. Both models are evaluated on two datasets with different characteristics: the OCID-Grasp dataset, which contains real cluttered RGB-D scenes, and the single-object Cornell GraspDataset. Beyond reporting final accuracy, this work measures inference latency and memory across input resolutions, analyses the training behaviour and loss progression of both models, and quantifies how sensitive their results are to factors that are usually left unreported. Specifically, the Cornell evaluation is repeated across multiple seeds, and one configuration is repeated three times, to establish the variance of the result. The two internal hyperparameters of the TTT inner loop, the inner learning rate and the inner gradient scale, are varied to test whether the values tuned for image classification transfer to graspdetection. The efficiency difference is remarkable: at $1024 \times 1024$ with a batch of eight the TTT model runs $2.6$ times faster and uses $72\%$ less memory, with latency curves crossing at roughly 700 pixels for single images. The accuracy part is the opposite. The two models finish within half a point of each other at native resolution on OCID, while three identical Cornell runs span $13.5$ percentage points, which is more than any difference reported between the models on that dataset. Varying the TTT inner-loop constants changes no accuracy the experiment can detect, though the number of folds that fail to converge rises with both. Test-Time Training therefore changes how the cost of a grasp detection network scales, not how accurate it is.

14:35 - 14:55 Oelmaier, Manuel (BA Thesis)

Title: Hidden-State-Informed Token Eviction for Linear Attention Transformers

Advisor: Hu Cao

Keywords: Linear Attention; Transformer

Abstract: This thesis proposes Hidden-State-Informed Token Eviction (HSTE), which ranks tokens by the write residual of a delta-rule memory: tokens represented poorly by the recurrent state receive priority in the sparse cache. The score uses tensors already produced by the memory write and adds no parameters or matrix multiplication. Concurrent work, HOLA, independently uses the write-weighted form of the same signal with a different backbone and fusion mechanism. On the single-seed Mechanistic Architecture Design suite, the write-weighted score obtains the highest mean among the tested score maps, while the unweighted residual only slightly exceeds random admission under matched cache and sink settings. That ordering does not transfer to 360–380M language models pretrained on 15B FineWeb-Edu tokens: neither tested HSTE score improves aggregate short-context quality over the parametermatched sliding-window hybrid, whose cache budget differs. The results support dynamic non-local cache allocation more clearly than residual-based token ranking.

14:55 - 15:15 Chen, Tingsheng (BA Thesis)

Title: Surprise-Driven Memory: Unifying Recurrent and Episodic States in Language Models

Advisor: Hu Cao

Keywords: Test-Time Training (TTT); Language Models

Abstract:  Recurrent linear-attention models avoid a growing key–value cache by compressing earlier tokens into a fixed-size state. Older tokens are no longer separately addressable once they leave the local window. This thesis adds episodic memory (EM) to a Test-Time Training (TTT) recurrent layer in the Memory as Gate (MAG) architecture. TTT-MAG+EM stores one pooled key–value entry per completed segment, which later queries can attend to. Experiments on the Mechanistic Architecture Design (MAD) benchmark and full-scale language-model tasks compare segmentation, pooling, and positional encoding choices. On MAD, the best uniform and adaptive variants reach 96.9% and 97.1%, respectively, on selective copying, compared with 86.7% for TTT-MAG. Uniform segmentation still gives the highest MAD average. Adaptive segmentation improves selective copying with attention-mean pooling but performs worse with mean and surprise-weighted mean pooling. Encoding token keys before pooling matches or outperforms after-pooling encoding on selective copying. At full scale, most evaluated before-pooling encodings slightly improve the language-model average over the corresponding model without positional encoding. Episodic memory is most useful for ordered long-range retrieval. Its modest language-model gains do not justify the added training cost.

15:15 - 15:40 Long, Qian (MA Thesis)

Title: Test Time Self Evolving Intelligent Agent

Advisor: Hu Cao

Keywords: Test-Time Training (TTT); Self Evolving

Abstract: 

15:40 - 16:05 Xiaokun Zhen (MA Thesis)

Title: Revisiting Swin-Unet for Medical Image Segmentation

Advisor: Hu Cao

Keywords: Swin-Unet, 3D segmentation

Abstract:  Transformer-based medical image segmentation requires robust feature representations and effective encoder-decoder information transfer. Existing window-based architectures, however, have been less extensively studied under changes in downstream resolution and attention-window size, while the use of multi-scale encoder features in Transformer decoders remains limited. This thesis addresses these issues by incorporating the attention design of Swin Transformer V2 into 2D Swin-Unet and investigating FuseUNet-inspired feature fusion in both 2D and 3D segmentation frameworks.

16:05 - 16:30 Qian, Yisong (MA Thesis)

Title: Optical Flow Guided 3D Occupancy Prediction for Dynamic Scenes

Advisor: Hu Cao

Keywords: Optical Flow; 3D Occupancy Prediction

Abstract:  Camera-based semantic scene completion (SSC) predicts voxel-wise geometry and semantics despite incomplete observations. This thesis uses left-view RGB images and precomputed stereo-depth guidance within VOXDET to study four problems: temporal misalignment, offset uncertainty, limited instance structure, and unequal regional evidence. Five extensionsare evaluated independently. Optical-flow-guided temporal regression fusion aligns historical image features and gates their contribution in voxel space. Probabilistic offset regression supplies uncertainty for inference-time aggregation. Instance-Aware Deformable Ray-19 Aggregation combines component-derived pseudo-instance supervision, center prediction, and spatial support features. Visibility bias favors estimated-visible support points for selected unobserved query regions. Pose-Aligned Masked Temporal Attention aligns independently lifted historical volumes and selects same-voxel evidence using depth-support reliability. On the SemanticKITTI validation split, the latter configuration with three causal historical frames achieves 49.16 SC IoU and 20.16 SSC mIoU, compared with 47.64 and 18.83 for the local VOXDET baseline. These single-seed results identify the strongest evaluated configuration while distinguishing geometric completion from semantic prediction.

16:30 - 16:55 Zhu, Hepeng (MA Thesis)

Title: Semantic-Driven and Feedback-Guided LLM Framework for Safety-Critical Scenario Generation

Advisor: Hu Cao

Keywords: Safety-Critical Scenario Generation; large language models (LLMs)

Abstract:  This thesis proposes a safety-critical scenario generation method targeting functional modules of autonomous driving systems. Based on real-world accident reports, it establishes a complete testing pipeline from high-value report screening and semantic scenario generation to closed-loop safetycritical scenario mutation. Building on previous work in which large language models (LLMs) were used to classify accident reports into perception, prediction, planning, and control categories, this thesis further combines rule-based scoring with LLM-based semantic scoring to select reports with high potential for scenario reproduction from each category. The selected accident reports are then transformed into structured semantic seeds and used to generate executable CARLA test scenarios under template constraints. Based on these scenarios, Responsibility-Sensitive Safety (RSS) and Spatio-Temporal Conflict (STC) metrics are employed to evaluate scenario risk and vehicle interactions. Together with collision root-cause analysis, these metrics guide subsequent scenario mutations, forming a closed-loop fuzzing method targeting specific functional modules. When the search stagnates, becomes repetitive, or deviates from the target module, an LLM is further employed to provide behavior-level mutation suggestions. The proposed method is evaluated in the Pylot–CARLA environment and compared with three existing fuzzers. The experimental results show that the proposed method can generate highly hazardous test scenarios while achieving stronger target-module specificity than the baseline methods.

16:55 - 17:20 Minxuan He (MA Thesis)

Title: Sparse Semantic Structure in Vision–Language Embeddings: Concept Factorization and Dual-Modality Grounding

Advisor: Yingbai Hu

Keywords: Multimodal Representation, Sparse Dictionary Learning, CLIP, Sparse Semantic Manifold, Interpretability

Abstract:  This thesis proposes SpLiCE-Aggregate, a post-hoc, two-stage framework that factorizes frozen CLIP image embeddings into a sparse dictionary of latent concept directions, each grounded both textually (via SpLiCE) and visually (via most-activating images). Applied to OpenCLIP ViT-B/32 representations of COCO categories, the recovered directions extend beyond object identity to scene, activity, and co-occurrence concepts, organized into a five-way taxonomy and validated via CLIPScore, grounding-word overlap, and ablation studies. These results support a sparse, factorized "sparse semantic manifold" view of the representation space, with SpLiCE-Aggregate achieving the best alignment–diversity trade-off among compared configurations.

17:20 - 17:45 Liangkai Li (MA Thesis)

Title: An Empirical Study of the Latent Conditioning Interface in Dual-System Vision-and-Language Navigation

Advisor: Yingbai Hu

Keywords: Vision-and-Language Navigation, Dual-System Architecture, Latent Conditioning, Reinforcement Learning, Embodied AI

Abstract:  This work studies the latent conditioning interface in a dual-system Vision-and-Language Navigation framework, where a slow VLM planner guides a fast trajectory policy. We evaluate a Qwen3-VL backbone upgrade, reinforcement-learning post-training, and trajectory fusion, and analyze how these modifications affect closed-loop navigation. The results show that backbone improvements transfer effectively to navigation performance, while RL and fusion gains are more limited, and the latent representation is shown to be instruction-conditioned and causally used by the downstream policy. 

17:45 - 18:10 Tianyu Jiang (MA Thesis)

Title: Impact-Friendly Control Design with Acceleration-Level Quadratic Optimization under RCM Constraints

Advisor: Yingbai Hu

Keywords: RCM, Quadratic Programming, Redundant Manipulator, Impact Prediction, Contact Force Control

Abstract:  This thesis proposes an acceleration-level QP control framework for redundant manipulators that maintains RCM constraints while enabling tool-tip tracking and contact interaction. A phase-dependent architecture integrates impact prediction during pre-impact motion and admittance-based force regulation during contact. Simulation results demonstrate that admittance reduces force RMSE by 42.6%–53.7%, while impact prediction reduces the first-contact force peak by 18.4%–59.1%.

18:10 - 18:30 Ghorbal, Mohamed (BA Thesis)

Title: Zero-Trust-Inspired Multi-Layer Security Framework for Robotic Telepresence

Advisor: Yansong Wu

Keywords: 

Abstract:  Networked robotic telepresence introduces new security risks under strict real-time constraints. We present a zero-trust-inspired security framework (ZTSF) for a 1 kHz leader–follower Franka Research 3 system. ZTSF shifts latency-intensive operations to session establishment, while its steady-state phase uses lightweight RBAC, motion validation, Noise-based authenticated encryption, eBPF microsegmentation, and follower-side verification. Evaluation on a physical FR3 system and through adversarial testing shows a mean one-way end-to-end latency below 27 µs across the tested payload sizes, comparable to the evaluated communication-security baselines. Formal analyses using Verifpal, TLA+, and Tamarin examine confidentiality, session binding, fail-closed receiver behavior, and replay resistance. The results demonstrate that selected zero-trust principles can satisfy the tested 1 kHz constraints, although the findings remain specific to the evaluated platform and configuration.

The I6 defense day is held monthly, usually on the last Friday of each month. The standard formats of talks are:

More information on preparing presentations can be found in the Thesis Submission Guidelines.