Jinyu Zhang, Lijun Han, Feng Jian, Lingxi Zhang, Hesheng Wang
8 min
In shared control, mobile robots combine human cognitive strengths with robotic precision. However, relying on direct teleoperation is cognitively demanding for humans in complex or cluttered environments. While existing planning-level shared control methods exist, they often struggle to balance autonomous robot performance with the need to fluidly accommodate human intentions. The authors sought to bridge this gap by developing a framework that predicts and incorporates short-term human motion intentions into the robot's path-planning process.
The authors propose a framework that treats intention prediction and path replanning as a joint Markov Decision Process (MDP), solved using deep reinforcement learning. A key innovation is the representation of short-term human intent as an 'intention domain'—a cone-shaped region in space that defines the likely path for the near future. The system uses a Voronoi-based trajectory generation algorithm to synthesize diverse human-like paths for training in simulation, eliminating the need for extensive human demonstration data. During operation, if a human intervenes, the learned policy triggers a replanning event if the current path deviates from the predicted intention domain, updating the robot's trajectory to align with the human's immediate goals.
In both simulation and real-world user studies using TurtleBot platforms, the framework outperformed traditional direct control and heuristic shared-control baselines. In a search-and-rescue user study, the proposed method reduced human interaction time by over 50% compared to manual control and provided a more intuitive experience, as indicated by higher System Usability Scale (SUS) scores and lower perceived workload across mental, physical, and frustration subscales. The approach effectively maintained safety, achieving higher clearance distances from obstacles while maintaining or improving task completion times.
In mobile robot shared control, effectively understanding human motion intention is critical for seamless human-robot collaboration. This paper presents a novel shared control framework with planning-level human intention prediction. It allows the operator to modify the robot's desired path based on inferred short-term motion intention. To represent future motion intentions, we introduce the concept of an intention domain and use it as a constraint for intention-aware path replanning. The intention-domain prediction and path replanning problems are jointly formulated as a Markov Decision Process and solved through deep reinforcement learning. In addition, a Voronoi-based human trajectory generation algorithm is further developed to train the policy entirely in simulation without human participation or demonstration data. We validate our approach through simulation and real-world experiments. In the real-robot user study, we compared the proposed method with both direct human control and a heuristic shared-control baseline. The proposed method reduced human interaction time by 60.4% and increased trajectory clearance by 13.2% relative to direct human control, further reduced human interaction time by 10.1% relative to the heuristic shared-control baseline, and achieved the lowest mean task completion time among the three control conditions. Subjective evaluations also indicate improved usability and reduced perceived workload.
Sam: So the cone says where the operator might be going, without claiming to know the exact path. What determines its size?
Alex: A learned policy chooses its length and width, using nearby obstacles, the final goal, the existing route, and recent movement. Length sets how far ahead the prediction reaches. Width controls how tightly the region follows the operator’s recent direction.
Sam: A broad cone could contain plenty of plausible routes, including ones I don’t want. How does the system commit?
Alex: It uses conventional path planning inside that constraint. Among obstacle-free points on the cone’s far boundary, it selects the point closest to the final destination. It plans a segment there, then connects that segment back to the final goal.
Sam: Put that into the search task. The initial route misses a newly discovered marker, and I start steering toward it.
Alex: Your intervention supplies motion history that can support a route change. The robot constructs the predicted region and revises its path. Once you judge that path approaches the sub-goal, you can release control, and the robot follows the revised route.
Sam: But a small correction could trigger the same sequence. Where does the system learn restraint?
Alex: The policy also decides whether to replan at all. It uses reinforcement learning, meaning decisions improve through rewarded experience in simulation. Rewards favour paths close to the simulated human’s intended path, penalize invalid regions, and discourage overly wide cones and excessive replanning.
Sam: That makes the training human crucial. If no people supply demonstrations, whose intentions is the policy learning?
Alex: Synthetic ones. The authors build a Voronoi graph, a route network derived from spaces between obstacles. Different routes capture choices such as passing an obstacle on either side. They add locally sampled waypoints and generate feasible trajectories through them.
Sam: Does that include someone temporarily moving away from the final destination? Otherwise, searching for intermediate targets would look like an error.
Alex: It does. The generator includes loops to represent exploration away from the final goal. This gives varied training routes without human demonstrations, though their diversity does not establish that they cover real human behaviour.
Sam: Let’s separate the evidence. What did simulation show before the real users entered?
Alex: On a generated test set of two hundred fifty-six trajectories, the learned method better matched intended paths than periodically replanning straight toward the final goal. It also replanned less often. Human-in-the-loop simulations then compared it with direct control and a heuristic planning assistant.
Sam: The heuristic seems the sharper test. What did learning add beyond reacting quickly whenever the operator intervened?
Alex: The heuristic replanned whenever the human intervened and expanded the cone according to available space. It triggered many more route updates. In narrow passages, input noise could misdirect its cone, and unnecessary updates could overwrite a route already matching the operator’s intention.
Sam: So responsiveness itself can become extra work. What happened when actual participants used the physical robot?
Alex: Ten participants performed an indoor search-and-rescue task, locating and marking sub-goals before reaching the final destination. Each used direct control, the heuristic assistant, and the learned assistant. Direct control still included local obstacle-avoidance assistance; it lacked autonomous route planning.
Sam: That matters: the comparison is not simply assistance versus none. Did learning outperform the simpler planning assistant too?
Alex: It reduced interaction time by a further ten percent relative to the heuristic, a statistically significant difference. The large saving against direct control therefore includes the benefit of planning-level assistance generally. Learning supplied a smaller additional gain.
Sam: Less joystick time does not necessarily mean faster or safer task execution. What else moved?
Alex: Average obstacle clearance increased by about thirteen percent relative to direct control. Both planning assistants significantly improved clearance, but this measures distance from obstacles, not proof of safety in every setting. The learned method had the lowest mean completion time, without a statistically significant completion-time difference.
Sam: And did people feel less burdened, or is that inferred from releasing the joystick?
Alex: Compared with direct control, the learned assistant significantly reduced reported mental demand, effort, and frustration. Its usability score was highest, but that difference was not statistically significant after correction. Subjective workload did not significantly distinguish the learned and heuristic assistants.
Sam: How far would you trust the result beyond this room? The training and task both seem quite structured.
Alex: I’d call it a promising proof of concept. The physical study was small, involved young adults, and used a compact indoor environment. Direct control always came first, leaving a possible practice effect despite randomized ordering of the shared-control methods.
Sam: What limitation would a developer notice first when the robot misunderstands an intervention?
Alex: Recent movement must indicate a reasonably stable intended direction. Abrupt switches, reversing, or tentative exploration can produce unnecessary replanning. The authors also flag fixed local map size around large obstacles and delayed recognition of intention when entering narrow corridors.
Sam: Who should read the full document, and where would you send them first?
Alex: Researchers building shared control or intention prediction should start with Intention-aware Path Replanning. Then read Comparative Analysis with Heuristic Baselines and the real-robot user study, keeping its ordering and sample in mind. For everyone else: useful robot assistance means learning when to change the plan—and when to leave it alone.
Sam: Thanks for listening.