Citation Context

...nfidence where it is. 8. Suitability for collaborative multi-robot mapping Multi-robot collaboration is a topic that is currently gaining significant attention in the scientific community (see e.g., (=-=Matarić, 1997-=-, Parker, 1996)). A sub-problem of multi-robot collaboration is multi-robot map acquisition (Lopez & Sierra, 1997). In the most general problem, one would like to place robots at arbitrary locations i...

Citation Context

...and are subjects to local maxima. This article placed considerably little emphasis on multi-robot collaboration during mapping. Clearly, many envisioned operational scenarios involved teams of robots =-=[63, 74, 90]-=-, and using multiple collaborative robot for building maps is clearly a worthwhile research goal. Existing techniques work well if the relative starting positions of all robots are known [12, 89], as ...

Citation Context

...ges the reward function from favoring easier behaviors to favoring more complex ones based on those easy behaviors. Balch [9] uses a shaped reinforcement reward function (earlier suggested by Mataric =-=[161]-=-) which depends on the number of partial steps fulfilled towards accomplishing the joint task. Balch shows that using a shaped reward leads to similar results to using a local reward, but in a signifi...

Citation Context

...cessfully applied to single-robot learning of various tasks, most commonly navigation [23], also hexapod walking [24], box-pushing [25], and has also been successfully applied to multi-robot learning =-=[26-=-]. Another relevant approach has been in teaching robots by demonstration, also referred to as imitation. [2] demonstrated simplied maze learning, i.e., learning turning behaviors, by following anothe...

Citation Context

...very little existing work on mathematical analysis of adaptive multirobot systems in dynamic environments. The few application-level studies to consider adaptation and learning in multi-robot systems =-=[13, 26, 23, 12, 4]-=-.shave not attempted to analyze collective behavior or performance of the system. To remedy this, we have recently extended [18] the stochastic processes framework developed in earlier works to a type...

Citation Context

...problems robots Preprint submitted to Elsevier Science 7 August 2006face when learning in the real-world, e.g., real-time learning in environments that are partially observable, dynamic, continuous (=-=Mataric, 1997-=-; Thrun and Mitchell, 1993; Thrun, 2002). However, learning quickly from interactions with a human teacher poses additional challenges (e.g., limited human patience, ambiguous human input, etc.) as we...

Citation Context

...ormance-based reinforcement: all agents are rewarded when any agent delivers an attractor. ffl Local shaped reinforcement: each agent is rewarded progressively as it accomplishes portions of the task =-=[Mat97]-=-. Full details on the formulation of these reward functions are provided in [Bal98]. 9.3 Foraging results: performance and diversity in simulation Performance in the multi-foraging task is measured as...

Citation Context

...ally to nonlinear and/or non-stationary tasks [58, 1]. As a result, the robot control program is generally more robust. Some of these approaches incorporate learning methods for control law synthesis =-=[47, 1, 35, 43, 45]-=-. If local control models are stable (in the sense of Lyapunov), then they can actively restrict the state to a neighborhood surrounding the attractor, thus approximately preserving some property in t...

Citation Context

... to speed up learning in such cases is to create a more informative reward signal using “shaping rewards” (Dorigo & Colombetti, 1998; Ng et al., 1999; Perkins & Hayes, 1996) or “progress indicators” (=-=Matarić, 1997-=-). Unfortunately, this requires significant design effort, results in less autonomous agents, and may alter the optimal solution, leading to unexpected behavior. We propose that agents that must repea...