Some Modifications to Our End-to-End UAV Planner

The one-stage planner YOPO maps a single depth image and the robot state directly to a set of candidate trajectories, trained by backpropagating through differentiable trajectory costs. This yields dense, geometrically informative supervision, but inherits the pathologies of soft-constrained optimization: the safety cost competes with the smoothness and goal-reaching terms, is non-convex across homotopy classes, and the single-piece polynomial is limited in expressiveness. In this report, we summarize several effective modifications. We adopt a two-piece MINCO parameterization, trading time for smoothness without altering the trajectory's spatial profile. We further lift YOPO's multi-modal prediction to span distinct homotopy classes, treating each motion primitive as a homotopy anchor that confines the trajectory to a feasible basin - without explicit safe-flight-corridor construction or front-end search. For dynamic feasibility, we impose barrier penalties on velocity and acceleration together with a curvature-dependent speed limit whose gradient acts only on the velocity, producing an adaptive-speed behavior that decelerates in cluttered regions or sharp turns. We replace score regression with a ranking loss, preventing small score errors from reordering the candidate set. These yield richer trajectory representations, safer obstacle avoidance, and more direct flight paths.

Beyond Monotonic Progress: Retry-Supervised Value Learning for Robot Imitation

Human demonstrations for robot imitation learning often contain mistakes and corrective behaviors, such as imprecise grasps, object misalignment, unstable contact, and repeated attempts. While these segments are commonly treated as noisy or suboptimal data, they provide valuable evidence about when execution deviates from a desirable path and how task feasibility can be restored. However, existing reward and value models often rely on monotonic progress assumptions, which capture coarse task advancement but may overlook local execution errors and corrective behaviors in imperfect demonstrations. In this work, we propose ReTVL (ReTry-Supervised Value Learning), a framework for learning mistake-sensitive value functions from mixed-quality robot demonstrations by leveraging retry events as sparse supervision. ReTVL captures the local degradation-and-recovery structure around mistakes by combining global progress calibration with local pairwise preference learning induced by sparsely annotated retry keypoints. The learned value model is then used to reweight demonstration chunks for downstream behavior cloning, reducing the influence of harmful execution errors while preserving useful corrective behaviors. Experiments on real-robot manipulation tasks show that ReTVL produces more fine-grained value estimates than progress-based baselines and improves imitation learning from imperfect demonstrations.

Reinforcing VLAs in Task-Agnostic World Models

Post-training Vision-Language-Action (VLA) models via reinforcement learning (RL) in learned world models has emerged as an effective strategy to adapt to new tasks without costly real-world interactions. However, while using imagined trajectories reduces the sample complexity of policy training, existing methods still heavily rely on task-specific data to fine-tune both the world and reward models, fundamentally limiting their scalability to unseen tasks. To overcome this, we argue that world and reward models should capture transferable physical priors that enable zero-shot inference. We propose RAW-Dream (Reinforcing VLAs in task-Agnostic World Dreams), a new paradigm that completely disentangles world model learning from downstream task dependencies. RAW-Dream utilizes a world model pre-trained on diverse task-free behaviors for predicting future rollouts, and an off-the-shelf Vision-Language Model (VLM) for reward generation. Because both components are task-agnostic, VLAs can be readily finetuned for any new task entirely within this zero-shot imagination. Furthermore, to mitigate world model hallucinations, we introduce a dual-noise verification mechanism to filter out unreliable rollouts. Extensive experiments across simulation and real-world settings demonstrate consistent performance gains, proving that generalized physical priors can effectively substitute for costly task-dependent data, offering a highly scalable roadmap for VLA adaptation.

Unified Noise Steering for Efficient Human-Guided VLA Adaptation

Diffusion-based vision-language-action (VLA) models have emerged as strong priors for robotic manipulation, yet adapting them to real-world distributions remains challenging. In particular, on-robot reinforcement learning (RL) is expensive and time-consuming, so effective adaptation depends on efficient policy improvement within a limited budget of real-world interactions. Noise-space RL lowers the cost by keeping the pretrained VLA fixed as a denoising generator while updating only a lightweight actor that predicts the noise. However, its performance is still limited due to inefficient autonomous exploration. Human corrective interventions can reduce this exploration burden, but they are naturally provided in action space, whereas noise-space finetuning requires supervision over noise variables. To address these challenges, we propose UniSteer, a Unified Noise Steering framework that combines human corrective guidance with noise-space RL through approximate action-to-noise inversion. Given a human corrective action, UniSteer inverts the frozen flow-matching decoder to recover a noise target, which provides supervised guidance for the same noise actor that is simultaneously optimized via reinforcement learning. Real-world experiments on diverse manipulation tasks show that UniSteer adapts more efficiently than strong noise-space RL and action-space human-in-the-loop baselines, improving the success rate from 20% to 90% in 66 minutes on average across four real-world adaptation tasks.

VLA-GSE: Boosting Parameter-Efficient Fine-Tuning in VLA with Generalized and Specialized Experts

Vision-language-action (VLA) models inherit rich visual-semantic priors from pre-trained vision-language backbones, but adapting them to robotic control remains challenging. Full fine-tuning (FFT) is prone to overfitting on downstream robotic data and catastrophic forgetting of pretrained vision-language capabilities. Parameter-efficient fine-tuning (PEFT) better preserves pre-trained knowledge, yet existing PEFT methods still struggle to adapt effectively to robot control tasks. To address this gap, we propose VLA-GSE, a parameter-efficient VLA fine-tuning framework that improves control adaptation while retaining PEFT's knowledge preservation advantage. Specifically, VLA-GSE (Generalized and Specialized Experts) is initialized by spectrally decomposing the frozen backbone, assigning leading singular components to generalized experts (shared experts) and disjoint residual components to specialized experts (routed experts). This decomposition improves adaptation capacity under a fixed trainable-parameter budget. Under a comparable parameter budget, VLA-GSE updates only 2.51% of the full model parameters and consistently outperforms strong FFT and PEFT baselines. It achieves 81.2% average zero-shot success on LIBERO-Plus, preserves pre-trained VLM capability comparably to LoRA on multimodal understanding benchmarks, and improves real-world manipulation success under multiple distribution shifts. Code is available at: https://github.com/YuhuaJiang2002/VLA-GSE

ParaMaP: Parallel Mapping and Collision-free Motion Planning for Reactive Robot Manipulation

Real-time and collision-free motion planning remains challenging for robotic manipulation in unknown environments due to continuous perception updates and the need for frequent online replanning. To address these challenges, we propose a parallel mapping and motion planning framework that tightly integrates Euclidean Distance Transform (EDT)-based environment representation with a sampling-based model predictive control (SMPC) planner. On the mapping side, a dense distance-field-based representation is constructed using a GPU-based EDT and augmented with a robot-masked update mechanism to prevent false self-collision detections during online perception. On the planning side, motion generation is formulated as a stochastic optimization problem with a unified objective function and efficiently solved by evaluating large batches of candidate rollouts in parallel within a SMPC framework, in which a geometrically consistent pose tracking metric defined on SE(3) is incorporated to ensure fast and accurate convergence to the target pose. The entire mapping and planning pipeline is implemented on the GPU to support high-frequency replanning. The effectiveness of the proposed framework is validated through extensive simulations and real-world experiments on a 7-DoF robotic manipulator. More details are available at: https://zxw610.github.io/ParaMaP.

YOPO-Rally: A Sim-to-Real Single-Stage Planner for Off-Road Terrain

Off-road navigation remains challenging for autonomous robots due to the harsh terrain and clustered obstacles. In this letter, we extend the YOPO (You Only Plan Once) end-to-end navigation framework to off-road environments, explicitly focusing on forest terrains, consisting of a high-performance, multi-sensor supported off-road simulator YOPO-Sim, a zero-shot transfer sim-to-real planner YOPO-Rally, and an MPC controller. Built on the Unity engine, the simulator can generate randomized forest environments and export depth images and point cloud maps for expert demonstrations, providing competitive performance with mainstream simulators. Terrain Traversability Analysis (TTA) processes cost maps, generating expert trajectories represented as non-uniform cubic Hermite curves. The planner integrates TTA and the pathfinding into a single neural network that inputs the depth image, current velocity, and the goal vector, and outputs multiple trajectory candidates with costs. The planner is trained by behavior cloning in the simulator and deployed directly into the real-world without fine-tuning. Finally, a series of simulated and real-world experiments is conducted to validate the performance of the proposed framework.

YOPOv2-Tracker: An End-to-End Agile Tracking and Navigation Framework from Perception to Action

Traditional target tracking pipelines including detection, mapping, navigation, and control are comprehensive but introduce high latency, limitting the agility of quadrotors. On the contrary, we follow the design principle of "less is more", striving to simplify the process while maintaining effectiveness. In this work, we propose an end-to-end agile tracking and navigation framework for quadrotors that directly maps the sensory observations to control commands. Importantly, leveraging the multimodal nature of navigation and detection tasks, our network maintains interpretability by explicitly integrating the independent modules of the traditional pipeline, rather than a crude action regression. In detail, we adopt a set of motion primitives as anchors to cover the searching space regarding the feasible region and potential target. Then we reformulate the trajectory optimization as regression of primitive offsets and associated costs considering the safety, smoothness, and other metrics. For tracking task, the trajectories are expected to approach the target and additional objectness scores are predicted. Subsequently, the predictions, after compensation for the estimated lumped disturbance, are transformed into thrust and attitude as control commands for swift response. During training, we seamlessly integrate traditional motion planning with deep learning by directly back-propagating the gradients of trajectory costs to the network, eliminating the need for expert demonstration in imitation learning and providing more direct guidance than reinforcement learning. Finally, we deploy the algorithm on a compact quadrotor and conduct real-world validations in both forest and building environments to demonstrate the efficiency of the proposed method.

AdaptiveStep: Automatically Dividing Reasoning Step through Model Confidence

Current approaches for training Process Reward Models (PRMs) often involve breaking down responses into multiple reasoning steps using rule-based techniques, such as using predefined placeholder tokens or setting the reasoning step's length into a fixed size. These approaches overlook the fact that specific words do not typically mark true decision points in a text. To address this, we propose AdaptiveStep, a method that divides reasoning steps based on the model's confidence in predicting the next word. This division method provides more decision-making information at each step, enhancing downstream tasks, such as reward model learning. Moreover, our method does not require manual annotation. We demonstrate its effectiveness through experiments with AdaptiveStep-trained PRMs in mathematical reasoning and code generation tasks. Experimental results indicate that the outcome PRM achieves state-of-the-art Best-of-N performance, surpassing greedy search strategy with token-level value-guided decoding, while also reducing construction costs by over 30% compared to existing open-source PRMs. In addition, we provide a thorough analysis and case study on the PRM's performance, transferability, and generalization capabilities.

Sampling-Based Hierarchical Trajectory Planning for Formation Flight

Formation flight of unmanned aerial vehicles (UAVs) poses significant challenges in terms of safety and formation keeping, particularly in cluttered environments. However, existing methods often struggle to simultaneously satisfy these two critical requirements. To address this issue, this paper proposes a sampling-based trajectory planning method with a hierarchical structure for formation flight in dense obstacle environments. To ensure reliable local sensing information sharing among UAVs, each UAV generates a safe flight corridor (SFC), which is transmitted to the leader UAV. Subsequently, a sampling-based formation guidance path generation method is designed as the front-end strategy, steering the formation to fly in the desired shape safely with the formation connectivity provided by the SFCs. Furthermore, a model predictive path integral (MPPI) based distributed trajectory optimization method is developed as the back-end part, which ensures the smoothness, safety and dynamics feasibility of the executable trajectory. To validate the efficiency of the developed algorithm, comprehensive simulation comparisons are conducted. The supplementary simulation video can be seen at https://www.youtube.com/watch?v=xSxbUN0tn1M.