Safe Flow Mpc Eval

safe-flow-mpc-eval

SafeFlowMPC: Predictive and Safe Trajectory Planning for Robot Manipulators with Learning-based Policies — Oelerich et al. (2026) (arXiv:2602.12794, 2026)

What this evaluates

Evaluates a hybrid trajectory planning framework that fuses flow matching with model-predictive control. It probes the system's ability to generate adaptive, collision-free motions for a 7-DoF robot manipulator while strictly enforcing safety constraints in real-time across global planning, reactive replanning, and dynamic human-robot handover scenarios.

Datasets

• Custom Robot Manipulation Benchmarks (Exp 1-3) — total ?; splits: (unstated); repo https://github.com/TU-Wien-ACIN-CDS/SafeFlowMPC

Metrics

• adherence to safety constraints (primary) — range: other
  • Comparative evaluation of trajectory quality, collision avoidance, and real-time execution at 10Hz against five baselines (VP-STO, BoundMPC, BC, FM, Ours with NL-Opt). Specific numerical metrics are not detailed in the provided excerpt.

Input / output format

Input: Current robot state, previous trajectory (for behavior cloning baseline), and environmental obstacle/target configurations.

Output: Sequential trajectory points or control commands for the 7-DoF KUKA manipulator.

Scoring recipe

for env in [global_context, obstructed_replanning, human_handover]:
    for method in [VP_STO, BoundMPC, BC, FM, SafeFlowMPC]:
        trajectory = method.plan(state=env.current_state, target=env.target)
        safety_violated = check_safety_constraints(trajectory)
        success = trajectory_reaches_target(trajectory) and not safety_violated
        record(method, env, success, safety_violated, execution_time)

Common pitfalls

• The excerpt does not specify exact quantitative metrics (e.g., success rate, collision count, or constraint violation magnitude), making direct metric reproduction impossible without the full paper.
• Real-time performance is tied to a fixed 10Hz execution frequency and specific solver configurations (e.g., IPOPT vs. real-time iteration), which heavily influence results.
• The human-robot handover experiment depends on an external dataset [7] that is not provided or described in this section.

Evidence (verbatim from paper)

Three experiments are presented in this section to evaluate performance, versatility, and adherence to safety constraints of the developed method. The first experiment utilizes our method to learn from a global trajectory planner to achieve fast local planning in a global context. The second experiment focuses on reactive online replanning in an obstructed environment. The third experiment is a dynamic human-robot object handover where the robot learns behavior from a human-human handover dataset [7].

Citation

@misc{oelerich2026safeflowmpc,
  title={SafeFlowMPC: Predictive and Safe Trajectory Planning for Robot Manipulators with Learning-based Policies},
  author={Oelerich et al. (2026)},
  year={2026},
  note={arXiv:2602.12794}
}

• arXiv: 2602.12794