Robot Learning & Embodied Control
Closed-Loop Trajectory Generation for Robot Manipulation Without External Motion Planners
Generated automatically from the limitations stated in 4 papers (CVPR, ICLR, NeurIPS), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Current keyframe-based robot manipulation policies predict static target end-effector poses and delegate trajectory generation to external sampling-based motion planners such as BiRRT. This decouples visual perception from execution dynamics, preventing real-time reactive corrections when environments change or objects slip mid-motion. Furthermore, relying on external planners introduces high computational latency and frequent planning failures in cluttered scenes where feasible kinematic paths are narrow.
Why it matters
Manipulation policies that continuously adapt their full arm trajectories to dynamic disturbances and moving objects in real time, eliminating the runtime latency and failure modes of external sampling-based planners.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Train a diffusion policy or action-chunking policy directly on continuous joint-space or Cartesian trajectory demonstrations in simulation (e.g., Robomimic or RLBench), measuring task success rate, execution time, and cycle-to-cycle replanning frequency under dynamic obstacle perturbations.
- 2
Formulate a hybrid policy that predicts dense waypoints alongside local collision-avoidance potential fields, evaluated against MoveIt/BiRRT baselines on trajectory smoothness, planning latency, and collision rates across cluttered tabletop manipulation environments.
- 3
Distill pre-computed BiRRT and trajectory optimization trajectories into a closed-loop reactive policy operating at 20–50 Hz from visual observations, measuring real-time tracking error and recovery rates under sudden target displacements.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
High-frequency visuomotor policies (e.g., standard Diffusion Policy or ACT) may already solve this for standard tabletop setups, leaving the open challenge restricted strictly to high-dimensional obstacle avoidance where motion planners maintain formal safety and collision-free guarantees that learned policies cannot yet match.
Evidence
Each paper's own statement of the limitation, verbatim.
- MoMaGen: Generating Demonstrations under Soft and Hard Constraints for Multi-Step Bimanual Mobile ManipulationICLR 2026
Relies on GPU-accelerated motion generators and is computationally intensive; simulation execution dominates cost (e.g., 100s to execute a base motion whose planning takes 18s) and base sampling has high variance when few poses are feasible
- PDFactor: Learning Tri-Perspective View Policy Diffusion Field for Multi-Task Robotic ManipulationCVPR 2025
Requires calibrated RGB-D multi-view setups and a motion planner (MoveIt! BiRRT) to execute predicted keyposes, so it is limited to keyframe-based manipulation, not closed-loop trajectory control
- FlowRAM: Grounding Flow Matching Policy with Region-Aware Mamba Framework for Robotic ManipulationCVPR 2025
Requires an external motion planner (BiRRT) to execute trajectories between predicted poses
- DynaRend: Learning 3D Dynamics via Masked Future Rendering for Robotic ManipulationNeurIPS 2025
Relies on an external low-level motion planner to execute predicted keyframe poses rather than producing action sequences end-to-end, limiting closed-loop control
Nearest existing work
- Speedup Patch: Learning a Plug-and-Play Policy to Accelerate Embodied ManipulationICML 2026
- DemoGrasp: Universal Dexterous Grasping from a Single DemonstrationICLR 2026
- Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic ManipulationCVPR 2025
- Scalable Trajectory Generation for Whole-Body Mobile ManipulationCVPR 2026
- Scaffolding Dexterous Manipulation with Vision-Language ModelsNeurIPS 2025
- PALM: Progress-Aware Policy Learning via Affordance Reasoning for Long-Horizon Robotic ManipulationCVPR 2026
- ManiSoft: Towards Vision-Language Manipulation for Soft Continuum RoboticsICML 2026
- Demystifying Action Space Design for Robotic Manipulation PoliciesICML 2026
- Action-Sketcher: From Reasoning to Action via Visual Sketches for Robotic ManipulationCVPR 2026
- RoboFlow4D: A Lightweight Flow World Model Toward Real-Time Flow-Guided Robotic ManipulationICML 2026
- Action-Geometry Prediction with 3D Geometric Prior for Bimanual ManipulationCVPR 2026
- Motion Synthesis with Sparse and Flexible Keyjoint ControlICCV 2025
- NeRF in the Palm of Your Hand: Corrective Augmentation for Robotics via Novel-View SynthesisCVPR 2023
- STEP: Warm-Started Visuomotor Policies with Spatiotemporal Consistency PredictionICML 2026
- Time Optimal Execution of Action Chunk Policies Beyond Demonstration SpeedICLR 2026