• Main.WorkshopICRA2014

Kod*lab Menu

Internal Links (Login Required)

Opportunities and Challenges of Joint Inference and Control in Mobile Robotics

Full-day workshop on May 31, 2014 (Sat) at ICRA 2014.

Workshop materials


Progress in robotics has yielded physical platforms that are slowly becoming more competent in the unstructured physical world. These growing capabilities raise the prospects for realizing their huge potential value as autonomous sensors: for survivors in search and rescue settings; for anomalies or threats in surveillance applications; for novelty or scientifically motivated collection in extraterrestrial exploration. At the same time, sensor technology has taken off, and the materials, communications and computational technology underlying its advance now raise the prospect of data torrents so vast that they cannot even be reasonably stored, much less processed and interpreted without some active, real- time interpretive control. Whether in electromagnetic ranging, electro-optical (laser , cameras), or, chemotaxis modalities, the multiplicity of tuning affordances and resulting highly variable focus of attention invites and even demands that algorithms for autonomous sensor management move into the realm of real time feedback control. Addressing such control problems raises novel questions of how to formulate tasks whose goals have as much to do with the agent’s state of information as with its material situation; how to couple internal variables such as belief state with physical degrees of freedom; and how to develop new representations that facilitate that integration and promote the expression of information-sensitive mechanical goals. This day-long workshop will sample the range of new opportunities, questions, and issues that arise as sensors become robots, and robots become sensors. New sensing modalities such as chemo-sensitive nanoscale devices raise the prospect of unparalleled access to perceptual domains long the unique province of animals: do we know how to use them? Traditionally “high-end” modalities such as radar have been transformed – both regarding cost (in footprint and dollars) as well as realtime tunability – by the advance of electronics and computation: can the sophisticated offline designs that emerged over nearly a century of waveform and receiver engineering be adapted for closed loop operation on mobile robots? Decades following the initial push for active vision in robotics, what is today’s state of the art, what theoretical insights have emerged, with what implications for practice, and how close to realtime implementation? Even assuming a nicely adaptive and computationally tractable sensorium, how should information-sensitive tasks be formulated to express the appropriate tradeoff between exploration and exploitation? How should strategic operation shift this tradeoff in the face of adversarial environments? How does a “distributed body” enhance or complicate the opportunities for joint inference and control over the sensorium? How does an imperfectly actuated body subject to a highly irregular, unpredictable environment support and benefit from the tunable sensorium?

Intended Audience and Format

We target robotics researchers working in the traditional area of active sensing as well as experts in technology and policy seeking to understand emerging opportunities for multidisciplinary advances bearing upon robotics. There has been a great deal of interest in this topic arising from various research communities and so we have a very full day of speakers planned. The format for the workshop would be roughly one dozen 20 min individual talks (e.g., two 1.5 hr sessions in the morning and afternoon respectively) followed by a panel discussion with audience participation at the end of the day.

Tentative Schedule

Time Speaker Topic
9 - 9:30 am Registration and opening remarks
9:30 - 10 am Daniel Lee (University of Pennsylvania) Dynamic belief states and information-theoretic decision making in adversarial environments
10 - 10:30 am Fabio Ramos (University of Sydney) Bayesian path planning for learning spatio-temporal processes
10:30 - 11 am Stefan Williams (ACFR, Sydney) TBA
11 - 11:30 am Break
11:30 am - 12 pm Stefano Soatto (UCLA) Active inference of representations: control’s role in visual perception and vice versa
12 - 12:30 pm Andrea Censi (MIT) Designing efficient low-latency sensorimotor control
12:30 - 2 pm Lunch
2 - 2:30 pm Phil Dames (University of Pennsylvania) Information-based multi-target localization using small teams of mobile sensors
2:30 - 3 pm Avik De (University of Pennsylvania) Toward dynamical sensor management: wall-following on RHex
3 - 3:30 pm Sertac Karaman (MIT) The impact of perception capabilities on agile robot motion: a statistical mechanics perspective
3:30 - 4 pm Break
4 - 5 pm Panel discussion


Copyright Kodlab, 2017