Swann robotics is a new approach to coordinate the behaviours of large number of relatively simple robots in decentralised manner. As the robots in the swann have only local perception and very limited local communication abilities, one of the challenges in designing swann robotic systems with desired collective behaviour is to understand the effect of individual behaviour on the group performance. This thesis dedicates the research on design and optimisation of interaction rules for a group of foraging robots that try to achieve energy efficiency collectively. The investigation starts with designing a set of interaction rules for the individual robots, inspired from the widely observed self-organisation phenomenon in biological system, so as improve the energy efficiency at the group level. A threshold-based controller, using two internal time thresholds - resting time and searching time threshold, is introduced to regulate the behaviours for the robot in order to improve the energy efficiency. Three cues: internal cues, social cues and environmental cues are then proposed to adjust the internal time thresholds in a self-organised manner. A number of strategies have been developed by combining these three cues and applied to the collective foraging task. Although the simulation results show that the robot swarm with adaptation mechanisms has the ability to guide the system towards energy optimisation collectively, there are difficulties in manually finding a set of parameters for the adaptation algorithm which can lead to the best energy efficiency under certain environmental conditions. This thesis focuses most of its effort into developing a macroscopic probabilistic model to understand the effect of individual parameters (internal time thresholds) on the performance of the system and therefore help to design the adaptation algorithm more efficiently. The modelling work is divided in two stages: A simplified situation for a swarm of homogeneous foraging robots without adaptation mechanism is considered first, then the macroscopic probabilistic model is extended for a robot swann with full adaptation ability. 3 The essential idea of the probabilistic modelling approach is to treat the interactions among robots, or between robots and environment, as stochastic events. First, a probabilistic finite state machine (PFSM), adapted from the robot controller, is used to describe the foraging task at the group level. A number of difference equations are then developed to capture the change of number of robot in each state. The state transition probabilities and other parameters used in the model are obtained through a novel geometrical approach, which makes sure that no free parameters exist in the model. In addition, the adaptation rules are encoded into the difference equations by introducing the concept of private resting/searching time thresholds and public resting/searching time thresholds. The proposed macroscopic model has been validated using simulation. The results show that the model achieves very good accuracy in predicting the net energy of the swarm, not only in the final stage but also in the instantaneous level. Finally, with the extended macroscopic model, a real-coded steady-state genetic algorithm (GA) is introduced to simplify the process of parameters selection for the adaptation algorithms. Experiments are carried out using the the best set of parameters found by the GA. It shows that the robot swarm with selected parameters can achieve a near-optimal energy efficiency under different environmental conditions.