Buoyant Choreographies: Harmonies of Light, Sound, and Human Connection
Dennis Hong, Yusuke Tanaka
cs.RO
2025-05-21
UCLA stages BALLU, a never-falling helium-balloon biped, as a fenceless show with fan walls, step-synced light and sound, and visitor control.
A walking robot in a gallery fails first on safety, not on gait quality. Conventional bipeds are heavy and top-heavy, so shows put up fences, tethers, or a deadman switch. Visitors watch. They do not share the floor.
Dennis Hong and Yusuke Tanaka at UCLA Mechanical and Aerospace Engineering turn an existing platform, BALLU (Buoyancy Assisted Lightweight Legged Unit), into an interactive installation. A helium balloon supplies near-neutral buoyancy, so the body is hard to tip over; articulated legs still walk and jump. The paper is not a new controller score. It stages the "does not fall" property as a fenceless, visitor-perturbable performance, layered on top of the permanent BALLU exhibit at the Seoul Robotics and Artificial Intelligence Museum (Seoul RAIM) with light, sound, and multi-robot staging.
BALLU's body is a helium balloon; the legs are articulated mechanisms. Each leg is its own robot: a microcontroller plus IMU, barometer, temperature, force/contact, and power sensors, talking to the other legs over both sub-GHz and 2.4 GHz radios to cut interference. Legs act as agents and then compose a whole-body gait. The paper calls this leg-as-a-robot. Cheap and light are the upsides. Buoyancy-assisted walking is highly nonlinear, so precise control is hard; the authors flag the platform as a fit for data-driven methods, not a finished position-servo demo.
The gallery has no physical fence. Four corner fans define an invisible motion volume; visitors may walk in. Extra upward fans lift the motion off the floor into 3D. At the end of each cycle, the fans swing toward the center and blow the robots into a gathering. Boundary, clustering, and lift are all done with airflow.
Light and sound are tied to the gait. RGB LEDs on the legs change color on every step and trigger musical notes, so walking becomes a glowing melody. Visitors get three entry points: a gamepad that scripts steps and rhythm; a light touch on the body; a redirected fan. Those disturbances change both gait and melody, so every run is one-off. Unpredictability is treated as a feature. The buoyancy system is already hard to servo tightly; the show lets people add noise on purpose.
This is a two-page exhibition note. There is no controlled experiment, no success rate, no pose error, no visitor survey. What the text and figures actually pin down:
| Item | What the paper states |
| Scale | Multiple robots walking at once; a figure shows at least five |
| Boundary | No physical fence; four corner fans form the arena |
| Cycle | Fans converge at cycle end and blow robots to the center |
| Interface | Gamepad, touch, and redirected airflow |
| Deployment | Permanent exhibit already at Seoul RAIM; this piece adds light, sound, and group staging |
| Prior work | BALLU2 in Frontiers in Robotics and AI 2021, framed as a safe, cheap buoyancy-assisted biped |
There is no quantitative comparison against any other exhibit robot or safe biped. The sensor list and control architecture are specific; the evaluation section is empty.
For embodied AI and human-robot coexistence, BALLU is a physics shortcut: buoyancy removes falling, so shared-floor interaction no longer depends on fences and e-stops. Independent legs, dual-band radio, and a full sensor suite mean this is not a balloon with sticks. It is still a distributed control problem. Strong nonlinearity and hard precision make it a sandbox for data-driven methods, more useful than a costume. If someone trains on this hardware, a fall is unlikely to wreck the next sample.
For museums, no fence plus visitor disturbance turns the audience into co-choreographers. Fans as walls and a center gather as narrative turn engineering constraints into staging. The helium balloon is both actuator and set piece.
This paper does not propose a new algorithm. It turns an existing platform into something a museum can put on the floor. The value is the setting, not a metric.
Almost nothing is measured. Helium leak, battery life, outdoor wind, and multi-robot packet loss, all of which a live show will hit, are unmentioned. Whether a fan boundary stays "invisible" in a dense crowd is shown only in renderings, with no traffic or collision counts. How gamepad, touch, and airflow are arbitrated when they conflict is unspecified.
"Never falls" is a buoyancy claim, not a control proof. A torn balloon, a helium shortfall, or an upward fan driving the body into an uncontrolled height are failure modes left out. The scientific challenge of leg-as-a-robot is named and then dropped; this paper has no control experiment, and the reader still does not know wireless sync latency or drop rate.
Read it as an exhibition tech note and the information is enough. Read it as a robotics algorithms paper and it is thin. Anyone chasing the controller should go back to BALLU2 (2021), not this document.