The goal of the project is to build an anthropomorphic two-armed robot together with a multi-camera vision system. The geometry of the system was inspired by the upper part of a human body, as shown in the pictures.

• The intelligent workbench
  The robot interacts with a human as an assistant in the workbench tasks. Such tasks could be for example the execution of chemical analysis, where the human decides what to do and the robot performs the frequent standard procedures or dangerous parts of the analysis. Based on simple commands like pick this or pour this into the robot adapts to become a useful helper for the human. Commands could be entered using gestures. The robot should use the vision system to locate and measure objects.
• The semi automomous robot
  As before, the robot teams up with a human. It performs simple manipulative tasks prescribed by the human operator, who can specify relatively complex tasks of pick-and-place nature. The robot solves the task by breaking it down into a number of individual pick-and-place steps autonomously. This property is useful if the operator and robot are separated by a distance and communication lines are unreliable or communication carries a time-delay. Another area where such autonomy is useful is when a robot arm and the visual system is mounted on a wheelchair and can be controlled by a handicapped person through use of a simple high-level interface.

• Manipulation using arms with a human-like geometry
• A Visual measurement and classification
• Control of a high degree of freedom system in a partially unpredictable, simply structured task environment.

The Manipulator Hardware

Our arm hardware, as shown in the laboratory overview, is based on two classically designed industrial manipulators. The motoric and sensory properties of such devices are naturally light-years away from their biological counterparts. The development of appropriate technical solutions for the mechanics and the tactile and visual sensors for a humanoid robot is still at a very early stage, and as long as this continues to be the case, with the components remaining as primitive as they currently are, there is little chance of building a system even mildly comparable to a biological system, let alone to a human being. Many other research teams are also working on this problem. Due to our manpower limitations and the lack of equipment for shaping metal ourselves, we have to accept this fact and focus on solving the resulting problems in vision and control using existing hardware.

The Vision System

The vision system provides the main sensory input to JANUS. It produces the following pieces of information for the control system:

• A scaled wire frame model of "interesting" objects, e.g. the objects that are to be grabbed.
• A rough location model of the "less interesting" objects, e.g. the objects which form obstacles in the movement of the arm.
• Classification of objects as instances of the base model or as "aliens".
• Location of pointer marks (e.g. a finger tip, or a laser pointer).
• Recognition of simple gestures (e.g. command pick or place).

The Control System

Currently most of our manpower has been invested in the development of the control system (see history of the project). The control system is based on a large set of local heuristics that form a community of reflective adaptive agents communicating using a blackboard-based logic.

The JANUS Laboratory

The lab contains the apparatus for experiments with the manipulators and the vision hardware. This includes a PC cluster which is used to run the arm control and vision software.

The Manipulators

The manipulators are mounted on a wall above a 2x2 m table. The experimenter can stand in front of the table and may reach into the cell to add or remove objects or to point at things. The following picture shows one of the manipulators in detail.

A few facts about the manipulators. manufactured by NIKO (a German company) 8 degrees of freedom outstretched length about 1.5 m maximal payload at hand is about 1 kg weight of the whole arm is about 40 kg driven by DC motors each joint contains an incremental angle decoder the maximal rotation speed of the joints is about 60 degrees per second

We chose this arm because, compared to more standard arms, the geometry fits quite well to human-like usage. Our kind of arm control is different to that used in industrial methods, therefore we were forced to build our own motoric control and amplifier hardware to ensure optimal controllablity of the manipulators.

A two finger gripper is mounted at the tool center plate of both manipulators. The gripper has the following charateristics.

manufactured by PRAEMETA (a German company) two parallel sliding fingers contact switches at the end of each finger jaw width from 5 cm to 15 cm DC motor driven incremental decoder at the motor weight about 400 g needs about 10 seconds for full close/open

We chose this kind of gripper because the DC motors allow continuous control of gripper jaw width. Additionally the pincer extendors allow a large stroke between open and close gripper. The contact switches allow the grasping of objects with a given force. A disadvantage is that the fingers require some time to close and the gripper itself is somewhat heavy.

Vision Hardware

The vision system consists of a stereo camera mounted in the "head" of JANUS and of a few (3-5) cameras mounted on tripods around the table.

 The "neck" is comprised of a basic two-joint manipulator. We use SONY color CCD cameras due to their low price. The cameras are connected to a Microtron MFG Inspecta frame grabber, which is able to grab frames with a few hundred Hz.

The low level motor driver

All motoric joints including the two manipulators, the grippers and the head are driven by DC motors with the same kind of interface. This enables us to build a standard amplifier and PID module for all kinds of motoric activities. At the moment we have in total 20 joints to control. As PID controllers we use MTC2 cards (manufactured by JANZ) mounted in two industrial 486/100 PCs running Linux. The amplifier module is a simple DC amplitude modulated amplifier. By avoiding a phase modulation amplifier, we reduce the AC components of the motor power supply to nearly zero, which gives us optimal smooth and stable motor behaviour. A drawback of our "old fashioned" way of control is that we need a little "cooling tower" to carry away the extra heat generated.

The PC Cluster

All arm control and vision algorithms of our system are based on many concurrently running local heuristics. These heuristics form agents that communicate via various blackboards. The philosophy of such algorithms allows a highly parallel model of execution. Our control algorithms use in all a few hundred independent processing units. Unfortunately we can't afford so many processors, therefore we use sequential scheduling that enables execution on a few CPUs. Initially the system was developed to run on a single SUN Sparc 20 computer. Meanwhile we have constructed a cluster of Pentium 166 PCs connected by a 100 Mb Ethernet. A few SUN workstations serve as terminals. The cluster computers run Linux as their operating system. The blackboards are mirrored across the cluster to enable the agents to communicate without knowing anything about the connectivity of the cluster. This method is used for both manipulator control and the vision system.