When several quadrotors carry one payload on cables, each drone controls the payload only through the direction and size of its thrust. Its heading, meaning its rotation about its own thrust axis, has no effect on the cable forces in the ideal model. This proposal treats that spare degree of freedom as a sensor. Each drone turns its body-fixed camera toward whatever the team most needs to see (the payload, a teammate, the path ahead), and the transport controller is written so that these turns cannot disturb the payload. Today's best multi-drone cable transport relies on an external motion-capture system, because onboard cameras have narrow fields of view and the drones hide one another. Turning the camera without disturbing the load is what could let such a team work from its own sensors.
This only works if the attitude controller keeps heading and thrust direction separate. Most rotation representations do not. Euler angles (roll, pitch, yaw) lose a degree of freedom when pitch reaches ±90°, the failure called gimbal lock, and they couple the axes well before that point. A standard controller on the full rotation group SO(3) lets a large yaw correction tilt the thrust vector for a moment. On a single drone that tilt causes drift. On a cable team the same transient becomes a jerk on a shared payload that every other drone feels. We propose to control each drone on S²×S¹, a thrust direction on the unit sphere times a heading on the circle, which removes this leak. The heading circle then becomes the action space for a decentralized problem of deciding who looks where, solved under partial observability.
Core Claim
In cooperative cable transport, a geometric attitude controller that splits each drone's orientation into thrust direction and heading makes camera pointing a disturbance-free action channel. Coordinating that channel across drones, together with the spare cable-tension directions available when there are three or more robots, lets a team estimate the payload's full position and orientation from onboard cameras alone. We expect payload tracking to approach motion-capture performance. We also expect the same heading schedule, run through a standard SO(3) controller, to measurably degrade payload tracking. That second prediction is the falsifiable part of the claim, and it isolates the geometric ingredient.
Why this is open
Three existing results each supply part of the idea, and none combines them. The first two motivate the problem.
Nonlinear Model Predictive Control for Cooperative Transportation and Manipulation of Cable Suspended Payloads with Multiple Quadrotors plans the payload's position and orientation directly on SE(3), the group of 3D positions and rotations. It uses a lightweight state for real-time optimization and exploits the null space of the tension distribution, meaning cable-force combinations that leave the payload motion unchanged, to keep robots apart. However, it takes every state from Vicon, collapses the drones into a lumped payload wrench, and leaves each drone's heading uncoordinated. PCMPC constrains one drone's full attitude so that its downward camera keeps a tagged slung load in view, using only a camera and an IMU. That single agent already uses 90% of the onboard CPU. Because it spends tilt, a degree of freedom that also moves the load, on perception, it cannot extend directly to a team whose tilts are pinned by cable-tension balance.
The third supplies the method. Trajectory Tracking for Quadrotors with Attitude Control on S2×S1 defines heading error by moving the reference heading along the shortest path on the sphere (parallel transport) from the desired thrust direction to the current one. It shows that this removes "control effort bleed", the lateral drift caused by 180° yaw maneuvers. Its evidence covers a single simulated drone with perfect state knowledge.
Nobody has pointed out that, in a coupled team, this decoupling turns heading into a free perception resource. That observation is the new element here.
Several further lines of work constrain the design. Nonlinear Robust Tracking Control of a Quadrotor UAV on SE(3) gives coordinate-free, almost-globally stable tracking but assumes full noise-free state feedback. We keep its position loop and its error functions and replace the attitude loop. Reinforcement Learning with Lie Group Orientations for Robotics shows that feeding and emitting rotations through the Lie-algebra log and exp maps speeds up learning relative to Euler angles or quaternions. The Lie algebra is the flat 3D tangent space of rotations, and log and exp map between it and the rotation group. That work is single-agent and fully observed, which suggests that any learned part of our system should see rotations through log coordinates and act on a well-behaved space. The circle S¹ qualifies trivially.
On the multi-agent side, COMRAD reports that centralized-critic methods can underperform fully independent learners on asymmetric-role tasks. It also reports that emergent coordination through physical signals remains unsolved. This is why we keep the learned component small and test independent against centralized training rather than assuming that centralized training helps.
Research Program
The system has three layers per drone and runs without a central computer during execution.
The bottom layer is a tracking controller on S²×S¹. It combines the Lee et al. position loop, which yields a desired force for the drone's cable, with the Kooijman et al. attitude error. That error regulates thrust direction on the sphere and heading, via parallel transport, on the circle. We will extend the single-drone decoupling argument to the cable-coupled model: with cables attached at each drone's center of mass, the heading subsystem enters the payload dynamics only through yaw reaction torque. We will bound that residual, so the claim "heading does not disturb the payload" becomes a stated inequality rather than a slogan.
The middle layer estimates the shared payload state. Each drone runs a filter on SE(3) for the payload pose, with errors expressed in log coordinates, so no Euler angles or quaternion sign ambiguities appear anywhere. It updates from its own tag detections when the payload is in view and from its own cable direction (a point on S²) inferred from IMU-measured tension. When bandwidth allows, drones exchange compact beliefs, which they fuse conservatively in the tangent space. Occlusions produce occasional wildly wrong rotation estimates, including 180° flips from symmetric tags. We will test whether a heavy-tailed measurement model inspired by A Laplace-inspired Distribution on SO(3) for Probabilistic Rotation Estimation resists these outliers better than a Gaussian model. That paper found that Gaussian-like rotation losses let such outliers dominate on static images. Its normalizer needs a large discrete sum, so we will use the distribution's shape only in the robust update weight.
The top layer is the heading scheduler: each drone chooses where its camera points. Our default is a decentralized greedy rule. Each drone turns toward the target whose observation most reduces the uncertainty of its local payload belief, discounted by what it expects teammates to cover. We compare this rule against a learned recurrent policy trained with independent and centralized critics. The action is a heading rate on S¹, and rotation inputs are passed through log coordinates. For teams of three or more, a second knob moves each drone within the tension null space toward positions that reduce mutual occlusion, reusing the ARPL null-space mechanism for visibility instead of separation. For this to matter, each drone carries a forward-tilted rather than downward camera; with a downward camera, heading would barely change what the drone sees. We state that dependence openly.
Baselines
The naive reference is fixed heading. Every drone holds a constant heading with the same onboard estimator and S²×S¹ controller and does no active pointing. This shows how much of the benefit comes from pointing at all, as opposed to having onboard estimation.
The strongest existing method is the ARPL cooperative NMPC. It reports 3.8 cm X/Y and 9.7 cm Z payload RMSE on a circular path with three Snapdragon quadrotors, 1 m cables, and a 232 g payload under Vicon feedback (Nonlinear Model Predictive Control for Cooperative Transportation and Manipulation of Cable Suspended Payloads with Multiple Quadrotors). We run it twice. Run with ground-truth state, it serves as the oracle ceiling. Run with our onboard estimator's output in place of Vicon, it serves as a fair competitor, since the planner is unchanged and only the information differs. Beating it means that our full system, with onboard sensing only, has lower payload RMSE than the NMPC fed by the same onboard sensors, on the same trajectories and with equal gain-tuning effort. It also means that we come within a stated factor of the Vicon ceiling; we set the target at twice the reported RMSE. No onboard-only multi-drone number exists on Lacuna, so that comparison has no prior figure.
A second competitor is PCMPC-style perception: each drone gets its own full-attitude field-of-view constraint and spends tilt, not heading, to keep the payload in view (PCMPC). The paper reports visual estimation error "within a few centimeters" at about 4 m/s for a single drone. No reported multi-drone result was found on Lacuna.
The key ablation is identical sensors, estimator, and heading schedule, with the attitude loop replaced by the Lee et al. SO(3) controller. An Euler-angle yaw-command variant appears as a secondary ablation.
An upper-bound reference uses an ideal massless camera gimbal with the same scheduler.
All controllers get the same tuning budget, a fixed number of gain-search trials on held-out trajectories. All learned schedulers get the same number of environment steps and seeds.
Why not just…?
Why not just keep the standard SO(3) controller and add yaw commands? This is the central threat, and the main ablation settles it. Kooijman et al. showed that large yaw maneuvers under SO(3) control cause lateral drift in a single drone. Our claim is that a cable team amplifies this, because a transient thrust tilt changes one cable's tension and so disturbs all the others. We will sweep the aggressiveness of the heading schedule. If payload RMSE under the SO(3) controller stays flat as heading rates increase, the geometric part of the proposal is unnecessary, and we will say so.
Why not just use Euler angles and keep pitch small? Near hover, pitch rarely nears 90°. Representation Issues in the ML Estimation of Camera Motion shows that minimal coordinates distort estimation landscapes well before the singularity, and Omnidirectional bipedal walking with direct fused angle feedback mechanisms shows that Euler yaw is entangled with roll and pitch. In our setting, the Euler yaw that the scheduler would command is not rotation about the thrust axis once the drone tilts, so commanded "pure pointing" leaks into thrust. The Euler variant in the ablation measures this directly. We expect a small gap in gentle flight and a larger one during gusts or tension transients, and we will report both regimes.
Why not just add a camera gimbal? A gimbal decouples the camera from the body in hardware, and our ideal-gimbal baseline may match or beat us. The case for the heading channel is that it costs no mass, no extra actuator, and no failure point on payload-limited vehicles. The honest question is how much of the ideal gimbal's benefit the free S¹ channel recovers. If it recovers little, the result is negative and informative.
Why not just train end-to-end multi-agent reinforcement learning with recurrent memory? Recurrent QMIX or MAPPO could in principle learn transport and perception jointly. However, COMRAD reports unsolved signaling tasks and harm from centralization in embodied settings. Learning thrust control on a cable-coupled, underactuated system also discards the stability guarantees that make hardware tests safe. We therefore restrict learning to the heading circle. If the greedy rule matches the learned scheduler, we will report the simpler rule as the method.
Proposed Experiments
All development runs in RotorTM, which models slack-to-taut cable transitions and includes SE(3) geometric controllers. We add rendered forward-tilted cameras, tag detection with occlusion by teammates and payload, IMU noise, and communication dropouts. RotorTM lacks motor dynamics and rotor drag, so we inject yaw-reaction torque and actuator saturation explicitly; this matters for the decoupling bound.
Tasks follow the ARPL protocol, a circular payload path with three drones, extended with figure-eights, 180° payload yaw turns, gust pulses that throw drones out of near-hover, and team sizes from three to five. Metrics are payload position and orientation RMSE (orientation as geodesic angle, never Euler differences), estimator error against simulator truth, fraction of time each drone and the team hold a payload observation, cable slack events, and minimum inter-drone separation.
The first decisive check is the heading-aggressiveness sweep described above, comparing S²×S¹ against SO(3) and Euler. The claim requires S²×S¹ payload error to stay nearly flat while the others rise.
The second check asks whether coordinated pointing beats fixed heading on onboard-only RMSE and closes a stated fraction of the gap to Vicon NMPC.
The third check compares greedy, independent-learned, and centrally trained schedulers at equal sample budgets. This tests whether the COMRAD centralization finding carries over to a physical coordination task.
The fourth check compares Gaussian against heavy-tailed rotation updates under induced tag flips.
If the simulation results hold, a three-drone hardware replication on an ARPL-like rig would use Vicon for ground truth only, never as controller input.
References
- Nonlinear Model Predictive Control for Cooperative Transportation and Manipulation of Cable Suspended Payloads with Multiple Quadrotors
- PCMPC: Perception-Constrained Model Predictive Control for Quadrotors with Suspended Loads using a Single Camera and IMU
- Trajectory Tracking for Quadrotors with Attitude Control on S2×S1
- Nonlinear robust tracking control of a quadrotor UAV on SE(3)
- Reinforcement Learning with Lie Group Orientations for Robotics
- COMRAD: A Benchmark for Embodied Cooperative Multi-Agent Reinforcement Learning
- A Laplace-inspired Distribution on SO(3) for Probabilistic Rotation Estimation
- Representation Issues in the ML Estimation of Camera Motion
- Omnidirectional bipedal walking with direct fused angle feedback mechanisms
- RotorTM: A Flexible Simulator for Aerial Transportation and Manipulation