Constraining Streaming Flow Models for Adapting Learned Robot Trajectory Distributions

TL;DR

Constraint-Aware Streaming Flow (CASF), a framework that augments streaming flow policies with constraint-dependent metrics that reshape the learned velocity field during execution, showing that it produces constraint-satisfying trajectories that remain smooth, feasible, and dynamically consistent, outperforming standard post-hoc projection baselines.

Abstract

Robot motion distributions often exhibit multi-modality and require flexible generative models for accurate representation. Streaming Flow Policies (SFPs) have recently emerged as a powerful paradigm for generating robot trajectories by integrating learned velocity fields directly in action space, enabling smooth and reactive control. However, existing formulations lack mechanisms for adapting trajectories post-training to enforce safety and task-specific constraints. We propose Constraint-Aware Streaming Flow (CASF), a framework that augments streaming flow policies with constraint-dependent metrics that reshape the learned velocity field during execution. CASF models each constraint, defined in either the robot's workspace or configuration space, as a differentiable distance function that is converted into a local metric and pulled back into the robot's control space. Far from restricted regions, the resulting metric reduces to the identity; near constraint boundaries, it smoothly attenuates or redirects motion, effectively deforming the underlying flow to maintain safety. This allows trajectories to be adapted in real time, ensuring that robot actions respect joint limits, avoid collisions, and remain within feasible workspaces, while preserving the multi-modal and reactive properties of streaming flow policies. We demonstrate CASF in simulated and real-world manipulation tasks, showing that it produces constraint-satisfying trajectories that remain smooth, feasible, and dynamically consistent, outperforming standard post-hoc projection baselines.

Figures (8)

• Figure 1: Overview of CASF and its effect. Left: collisions produced by the baseline SFP. Right: collision-avoiding behaviours under CASF. Middle: the proposed CASF framework, which introduces geometry- and distance-aware constraints as a post-training module.
• Figure 2: Visualisation of learned neural signed distance fields (Neural SDFs). Left: Robomimic simulation scene with a vase obstacle placed on the table. Middle: zero level-set surface reconstructions of the learned SDF from two viewpoints (camera-aligned and top-down). Right: corresponding 2D SDF slices.
• Figure 3: We sample body points on the robot. Left: The joints and links of the manipulator. Middle: dense surface sampling over the full arm mesh. Right: link-wise surface sampling, preserving local geometry per link and used in our pullback-based constraint formulation.
• Figure 4: Comparison of representative rollouts under different constraint-handling strategies: No Shaping(left), Hard-Projection (left-second), CBF (right-second), and CASF (right). Each panel reports the final task score and number of steps.
• Figure 5: Qualitative evaluation of reactive obstacle avoidance on four LASA tasks (Line, Worm, Z-shape, and W-shape). Each task compares the raw policy, Hard-Projection, CBF, and CASF. Grey streamlines visualize the induced velocity fields, while orange curves show rollout trajectories. The blue dashed curve denotes the unshaped predicted trajectory, and the red curve in the first column indicates the ground-truth demonstration. Transparent reddish overlays highlight the correction masks introduced by each shaping method.