The visual processing would be the hard part - and that's not an Arduino problem. If you're able to put a PC in your solution then PC + web cam plus application to to visual processing plus Arduino to drive the arm might be a viable solution. Depending on what object you're trying to recognise and what environment you're trying to recognise it in, the visual processing might be either trivially simple with OpenCV or it might be hideously difficult.