PROSPECT: Unified Streaming Vision-Language Navigation via Semantic--Spatial Fusion and Latent Predictive Representation

Multimodal large language models (MLLMs) have advanced zero-shot end-to-end Vision-Language Navigation (VLN), yet robust navigation requires not only semantic understanding but also predictive modeling of environment dynamics and spatial structure. We propose PROSPECT, a unified streaming navigation agent that couples a streaming Vision-Language-Action (VLA) policy with latent predictive representation learning. PROSPECT uses CUT3R as a streaming 3D foundation spatial encoder to produce long-context, absolute-scale spatial features, and fuses them with SigLIP semantic features via cross-attention. During training, we introduce learnable stream query tokens that query the streaming context and predict next-step 2D and 3D latent features (rather than pixels or explicit modalities), supervised in the latent spaces of frozen SigLIP and CUT3R teachers. The predictive branch shapes internal representations without inference overhead. Experiments on VLN-CE benchmarks and real-robot deployment demonstrate state-of-the-art performance and improved long-horizon robustness under diverse lighting. We will release code for the community soon.

Direction Matters: Learning Force Direction Enables Sim-to-Real Contact-Rich Manipulation

Sim-to-real transfer for contact-rich manipulation remains challenging due to the inherent discrepancy in contact dynamics. While existing methods often rely on costly real-world data or utilize blind compliance through fixed controllers, we propose a framework that leverages expert-designed controller logic for transfer. Inspired by the success of privileged supervision in kinematic tasks, we employ a human-designed finite state machine based position/force controller in simulation to provide privileged guidance. The resulting policy is trained to predict the end-effector pose, contact state, and crucially the desired contact force direction. Unlike force magnitudes, which are highly sensitive to simulation inaccuracies, force directions encode high-level task geometry and remain robust across the sim-to-real gap. At deployment, these predictions configure a force-aware admittance controller. By combining the policy's directional intent with a constant, low-cost manually tuned force magnitude, the system generates adaptive, task-aligned compliance. This tuning is lightweight, typically requiring only a single scalar per contact state. We provide theoretical analysis for stability and robustness to disturbances. Experiments on four real-world tasks, i.e., microwave opening, peg-in-hole, whiteboard wiping, and door opening, demonstrate that our approach significantly outperforms strong baselines in both success rate and robustness. Videos are available at: https://yifei-y.github.io/project-pages/DirectionMatters/.

Seeing to Act, Prompting to Specify: A Bayesian Factorization of Vision Language Action Policy

The pursuit of out-of-distribution generalization in Vision-Language-Action (VLA) models is often hindered by catastrophic forgetting of the Vision-Language Model (VLM) backbone during fine-tuning. While co-training with external reasoning data helps, it requires experienced tuning and data-related overhead. Beyond such external dependencies, we identify an intrinsic cause within VLA datasets: modality imbalance, where language diversity is much lower than visual and action diversity. This imbalance biases the model toward visual shortcuts and language forgetting. To address this, we introduce BayesVLA, a Bayesian factorization that decomposes the policy into a visual-action prior, supporting seeing-to-act, and a language-conditioned likelihood, enabling prompt-to-specify. This inherently preserves generalization and promotes instruction following. We further incorporate pre- and post-contact phases to better leverage pre-trained foundation models. Information-theoretic analysis formally validates our effectiveness in mitigating shortcut learning. Extensive experiments show superior generalization to unseen instructions, objects, and environments compared to existing methods. Project page is available at: https://xukechun.github.io/papers/BayesVLA.

Toward Embodiment Equivariant Vision-Language-Action Policy

Vision-language-action policies learn manipulation skills across tasks, environments and embodiments through large-scale pre-training. However, their ability to generalize to novel robot configurations remains limited. Most approaches emphasize model size, dataset scale and diversity while paying less attention to the design of action spaces. This leads to the configuration generalization problem, which requires costly adaptation. We address this challenge by formulating cross-embodiment pre-training as designing policies equivariant to embodiment configuration transformations. Building on this principle, we propose a framework that (i) establishes a embodiment equivariance theory for action space and policy design, (ii) introduces an action decoder that enforces configuration equivariance, and (iii) incorporates a geometry-aware network architecture to enhance embodiment-agnostic spatial reasoning. Extensive experiments in both simulation and real-world settings demonstrate that our approach improves pre-training effectiveness and enables efficient fine-tuning on novel robot embodiments. Our code is available at https://github.com/hhcaz/e2vla

TOP: Time Optimization Policy for Stable and Accurate Standing Manipulation with Humanoid Robots

Humanoid robots have the potential capability to perform a diverse range of manipulation tasks, but this is based on a robust and precise standing controller. Existing methods are either ill-suited to precisely control high-dimensional upper-body joints, or difficult to ensure both robustness and accuracy, especially when upper-body motions are fast. This paper proposes a novel time optimization policy (TOP), to train a standing manipulation control model that ensures balance, precision, and time efficiency simultaneously, with the idea of adjusting the time trajectory of upper-body motions but not only strengthening the disturbance resistance of the lower-body. Our approach consists of three parts. Firstly, we utilize motion prior to represent upper-body motions to enhance the coordination ability between the upper and lower-body by training a variational autoencoder (VAE). Then we decouple the whole-body control into an upper-body PD controller for precision and a lower-body RL controller to enhance robust stability. Finally, we train TOP method in conjunction with the decoupled controller and VAE to reduce the balance burden resulting from fast upper-body motions that would destabilize the robot and exceed the capabilities of the lower-body RL policy. The effectiveness of the proposed approach is evaluated via both simulation and real world experiments, which demonstrate the superiority on standing manipulation tasks stably and accurately. The project page can be found at https://anonymous.4open.science/w/top-258F/.

Disambiguate Gripper State in Grasp-Based Tasks: Pseudo-Tactile as Feedback Enables Pure Simulation Learning

Grasp-based manipulation tasks are fundamental to robots interacting with their environments, yet gripper state ambiguity significantly reduces the robustness of imitation learning policies for these tasks. Data-driven solutions face the challenge of high real-world data costs, while simulation data, despite its low costs, is limited by the sim-to-real gap. We identify the root cause of gripper state ambiguity as the lack of tactile feedback. To address this, we propose a novel approach employing pseudo-tactile as feedback, inspired by the idea of using a force-controlled gripper as a tactile sensor. This method enhances policy robustness without additional data collection and hardware involvement, while providing a noise-free binary gripper state observation for the policy and thus facilitating pure simulation learning to unleash the power of simulation. Experimental results across three real-world grasp-based tasks demonstrate the necessity, effectiveness, and efficiency of our approach.

Efficient Alignment of Unconditioned Action Prior for Language-conditioned Pick and Place in Clutter

We study the task of language-conditioned pick and place in clutter, where a robot should grasp a target object in open clutter and move it to a specified place. Some approaches learn end-to-end policies with features from vision foundation models, requiring large datasets. Others combine foundation models in a zero-shot setting, suffering from cascading errors. In addition, they primarily leverage vision and language foundation models, focusing less on action priors. In this paper, we aim to develop an effective policy by integrating foundation priors from vision, language, and action. We propose A$^2$, an action prior alignment method that aligns unconditioned action priors with 3D vision-language priors by learning one attention layer. The alignment formulation enables our policy to train with less data and preserve zero-shot generalization capabilities. We show that a shared policy for both pick and place actions enhances the performance for each task, and introduce a policy adaptation scheme to accommodate the multi-modal nature of actions. Extensive experiments in simulation and the real-world show that our policy achieves higher task success rates with fewer steps for both pick and place tasks in clutter, effectively generalizing to unseen objects and language instructions.

LESA: Learnable LLM Layer Scaling-Up

Training Large Language Models (LLMs) from scratch requires immense computational resources, making it prohibitively expensive. Model scaling-up offers a promising solution by leveraging the parameters of smaller models to create larger ones. However, existing depth scaling-up methods rely on empirical heuristic rules for layer duplication, which result in poorer initialization and slower convergence during continual pre-training. We propose \textbf{LESA}, a novel learnable method for depth scaling-up. By concatenating parameters from each layer and applying Singular Value Decomposition, we uncover latent patterns between layers, suggesting that inter-layer parameters can be learned. LESA uses a neural network to predict the parameters inserted between adjacent layers, enabling better initialization and faster training. Experiments show that LESA outperforms existing baselines, achieving superior performance with less than half the computational cost during continual pre-training. Extensive analyses demonstrate its effectiveness across different model sizes and tasks.

KVSharer: Efficient Inference via Layer-Wise Dissimilar KV Cache Sharing

The development of large language models (LLMs) has significantly expanded model sizes, resulting in substantial GPU memory requirements during inference. The key and value storage of the attention map in the KV (key-value) cache accounts for more than 80\% of this memory consumption. Nowadays, most existing KV cache compression methods focus on intra-layer compression within a single Transformer layer but few works consider layer-wise compression. In this paper, we propose a plug-and-play method called \textit{KVSharer}, which shares the KV cache between layers to achieve layer-wise compression. Rather than intuitively sharing based on higher similarity, we discover a counterintuitive phenomenon: sharing dissimilar KV caches better preserves the model performance. Experiments show that \textit{KVSharer} can reduce KV cache computation by 30\%, thereby lowering memory consumption without significantly impacting model performance and it can also achieve at least 1.3 times generation acceleration. Additionally, we verify that \textit{KVSharer} is compatible with existing intra-layer KV cache compression methods, and combining both can further save memory.

Lagrangian Motion Fields for Long-term Motion Generation

Long-term motion generation is a challenging task that requires producing coherent and realistic sequences over extended durations. Current methods primarily rely on framewise motion representations, which capture only static spatial details and overlook temporal dynamics. This approach leads to significant redundancy across the temporal dimension, complicating the generation of effective long-term motion. To overcome these limitations, we introduce the novel concept of Lagrangian Motion Fields, specifically designed for long-term motion generation. By treating each joint as a Lagrangian particle with uniform velocity over short intervals, our approach condenses motion representations into a series of "supermotions" (analogous to superpixels). This method seamlessly integrates static spatial information with interpretable temporal dynamics, transcending the limitations of existing network architectures and motion sequence content types. Our solution is versatile and lightweight, eliminating the need for neural network preprocessing. Our approach excels in tasks such as long-term music-to-dance generation and text-to-motion generation, offering enhanced efficiency, superior generation quality, and greater diversity compared to existing methods. Additionally, the adaptability of Lagrangian Motion Fields extends to applications like infinite motion looping and fine-grained controlled motion generation, highlighting its broad utility. Video demonstrations are available at \url{https://plyfager.github.io/LaMoG}.