Imitation learning with diffusion models has advanced robotic control by capturing multi-modal action distributions. However, existing approaches typically treat observations as high-level conditioning inputs to the denoising network, rather than integrating them into the stochastic dynamics of the diffusion process itself. As a result, sampling must begin from random Gaussian noise, weakening the coupling between perception and control and often yielding suboptimal performance. W e introduce Bridge-Policy, a generative visuomotor policy that explicitly embeds observations within the stochastic differential equation via a diffusion-bridge formulation. By constructing an observation-informed trajectory, BridgePolicy enables sampling to start from a rich, informative prior rather than random noise, substantially improving precision and reliability in control. A key challenge is that classical diffusion bridges connect distributions with matched dimensionality, whereas robotic observations are heterogeneous and multi-modal and do not naturally align with the action space. T o address this, we design a multi-modal fusion module and a semantic aligner that unify visual and state inputs and align observation and action representations, making the bridge applicable to heterogeneous robot data. Extensive experiments across 52 simulation tasks on three benchmarks and five real-world tasks demonstrate that BridgePolicy consistently outperforms state-of-the-art generative policies.

This paper introduces DiffTORI, which utilizes Diff erentiable T rajectory O ptimization as the policy representation to generate actions for deep R einforcement and I mitation learning. Trajectory optimization is a powerful and widely used algorithm in control, parameterized by a cost and a dynamics function.

Robot action primitives are agnostic to the exact geometry of the underlying robot, provided the robot is a manipulator arm. As noted in the related works section, Dynamic Motion Primitives (DMP) are an alternative skill formulation that is common robotics literature. Each primitive ran 200 low-level actions with a path length of five high level actions, while the low-level path length was 500. With raw actions, each episode took 16.49 We run an ablation to measure how often RAPS uses each primitive.

We present DEPS, an end-to-end algorithm for discovering parameterized skills from expert demonstrations. Our method learns parameterized skill policies jointly with a meta-policy that selects the appropriate discrete skill and continuous parameters at each timestep. Using a combination of temporal variational inference and information-theoretic regularization methods, we address the challenge of degeneracy common in latent variable models, ensuring that the learned skills are temporally extended, semantically meaningful, and adaptable. We empirically show that learning parameterized skills from multitask expert demonstrations significantly improves generalization to unseen tasks. Our method outperforms multitask as well as skill learning baselines on both LIBERO and MetaWorld benchmarks. We also demonstrate that DEPS discovers interpretable parameterized skills, such as an object grasping skill whose continuous arguments define the grasp location.

This paper introduces DiffTORI, which utilizes Diff erentiable T rajectory O ptimization as the policy representation to generate actions for deep R einforcement and I mitation learning. Trajectory optimization is a powerful and widely used algorithm in control, parameterized by a cost and a dynamics function.

The actor losses used in DoubleGum, SAC, and DDPG are all derived from the same principle. SAC (Haarnoja et al., 2018a,b) has a policy with learned variance and state-independent Section B.1 shows this for the actor losses of DoubleGum, SAC, and DDPG. We now relate the critic losses to each other, starting from the most general case, DoubleGum. The SAC noise model is derived from Equation 16 in three ways. In continuous control, Fujimoto et al. (2018) introduced Twin Networks, a method that improved Follow-up work selects a quantile estimate from an ensemble (Kuznetsov et al., 2020; Chen et al., 2021; Ball et al., 2023), which we demonstrate is Moskovitz et al. (2021) and Ball et al. (2023) showed that the appropriate Garg et al. (2023) present a method of estimating its value using Gumbel regression.

Preference-based reinforcement learning (PbRL) bypasses complex reward engineering by learning rewards directly from human preferences, enabling better alignment with human intentions. However, its effectiveness in multi-stage tasks, where agents sequentially perform sub-tasks (e.g., navigation, grasping), is limited by stage misalignment: Comparing segments from mismatched stages, such as movement versus manipulation, results in uninformative feedback, thus hindering policy learning. In this paper, we validate the stage misalignment issue through theoretical analysis and empirical experiments. To address this issue, we propose STage-AlIgned Reward learning (STAIR), which first learns a stage approximation based on temporal distance, then prioritizes comparisons within the same stage. Temporal distance is learned via contrastive learning, which groups temporally close states into coherent stages, without predefined task knowledge, and adapts dynamically to policy changes. Extensive experiments demonstrate STAIR's superiority in multi-stage tasks and competitive performance in single-stage tasks. Furthermore, human studies show that stages approximated by STAIR are consistent with human cognition, confirming its effectiveness in mitigating stage misalignment.