2. Biomimetic Robots for Robust Operation in Unstructured Environments University of California at Berkeley H. Kazerooni

3. The objective is to create robust and small walking machines. (Insects are good examples because they are small and robust.)

4. Observations from biological systems Engineering specifications that may provide some insights as how to make machines so their behavior is similar to our observations of biological systems.

5. Easy walk Requires little processing Fixed gait Utilizes pendulum like natural frequency to minimize energy One gets tired fast Lots of processing Terrain requires random gait Random gait will not allow any form of natural frequency -- thus walking is very inefficient Single D.O.F (Mechanism)Multiple D.O.F (Complex Machine) WALKING ROUGH TERRAINSMOOTH TERRAIN Walking Speed is independent of load -- i.e. students walk at the same speed regardless of their backpack load

6. You Stop Growing (L = Constant)

7. Single D.O.F (Mechanism)Multiple D.O.F (Complex Machine) Running ROUGH TERRAINSMOOTH TERRAIN

8. Full has shown that a substantial portion of locomotor control can reside in the mechanical design of the system and be simple: Biological Observations Control results from the properties of the parts and their morphological arrangement. Musculoskeletal units, leg segments and legs do much of the computations on their own by using segment mass, length, inertia, elasticity, and damping as “primitives”. Engineering Equivalence The system performance is function of the physical system; no feedback control has been used to alter the dynamics of the system.

9. Biological Observations During climbing, turning, and maneuvering over irregular terrain, animals use virtually the same gait as in horizontal locomotion - an alternating tripod. The animals appear to be playing the same feedforward program for running. There is no precise foot placement, no follow the leader gait, and a leg does not have time to react to tactile sensory feedback within a step. Engineering Equivalence A one degree of freedom system only. No need to design elaborate multi- variable robotic legs. Open loop within the system workspace

10. Biological Observations Position control using reflexes is improbable if not impossible The control algorithms are embedded in the form of the animal itself. The mechanical system - the morphology - can determine the extent of self-stabilization. In other words, there is no explicit feedback control of global variables such as hopping height, posture, or speed. The only control is local, at the joints. Engineering Equivalence No need for sensors for position speed, or force control The dynamics of the system is due to the dynamics of the hardware as designed by the designer with no alteration by feedback control. If there is no explicit feedback control of global variables such as hopping height, posture, or speed, therefore the control space (e.g. trajectory) must be limited.

11. The “Classical Robotics” technology may not be effective in design of small and robust walking machines

12. Design Specifications High Speed Mobility Small Size & Light Weight, (however constrained by fabrication technology, 3x4 inch body) No sensing, No feedback controls Cockroach foot path Compliance and Stability Simple Design (4 legs at this time) On-board power (Dictates the entire design Expected Speed (about 3”/sec)

13. Human foot Trajectory

14. Cockroach Foot Trajectory (Cutesy of Bob Full Lab)

15. Design of Mechanism to Mimic Cockroach Leg Trajectory

16. Path Requirements Slow gait on bottom, fast gait on top. Flat path at ground contact. Taller gait for high clearance. Longer gait for efficient walking. Robust geometry in the presence of fabrication inaccuracies.

17. Verification of Trajectories

18. Experimental Machine at UC-Berkeley size: 3.5"x3" speed: 3 inch/sec

20. Low-Level Control Fabrication High-Level Control MURI What passive properties are found in Nature? What properties in mechanical design? How should properties be varied for changing tasks, conditions ? Matching ideal impedance for unstructured dynamic tasks (Harvard) Guiding questions Preflexes: Muscle and Exoskeleton Impedance Measurements (Berkeley Bio.) Biological implications for Robotics Basic Compliant Mechanisms for Locomotion (Stanford) Variable compliance joints (Harvard, Stanford) Fast runner with biomimetic trajectory (Berkeley ME)

21. BioMimetic Robotics MURI Berkeley-Harvard Hopkins-Stanford MURI Low-Level Control Minimum Impedance Control Minimizing Interaction Forces in Exploration and Manipulation Jaydev P. Desai and Robert D. Howe Harvard University

22. Intrinsic Finger Stiffness vs. Force 4 subjects (Hajian and Howe 1997) 05101520 0 200 400 600 800 1000 1200 stiffness (N/m) finger tip force (N) Key robot capability for unstructured environments MURI Low-Level Control Impedance in Manipulation Example: Grasping in an unstructured environment –Object location uncertain. –Before contact: No interaction force. –Unexpected collision produces only small disturbance force f = k x if k is small.

23. MURI Low-Level Control Minimum Impedance Control for Grasping and Manipulation Goal: Build a simple robot gripper that can probe and grasp objects with minimum forces in unstructured environments. Approach: Combine biologically-inspired elements Low-impedance arm Minimum impedance controller Simple contact sensing

24. => Low impedance manipulator arm Year 1: Implemented testbed system, including hardware, software development system MURI Low-Level Control Variable Impedance Manipulation Testbed Whole-Arm Manipulator (Barrett Technology) Low moving mass Minimal friction Back driveable

25. MURI Low-Level Control How to Control Robot Motion with Low Stiffness? Conventional error-based position control law: Joint torque =  = K p (x d - x) + K d (v d - v) –Gain = stiffness: K p = (torque)/(position change) –Need high gain K p for small position error (x d - x) –If unexpected contact occurs => error (x d - x) becomes large => controller generates large force f ~ K p (x d - x).

26. MURI Low-Level Control Model-Based Position Control Law Joint torque =  = arm model + error terms Use arm model to generate feedforward torques that make robot follow desired trajectory:  (model) = arm dynamics + joint friction Arm model = –Dynamics - inertia, coriolis, gravity, etc. –Friction - each joint –Experimentally measured each term for the WAM arm testbed

27. MURI Low-Level Control Model-Based Position Control Law Joint torque =  = arm model + error terms Use error terms for minor corrections only  (error) = = K p (x p - x) + K d (v p - v) If model is accurate, low gains produce good control Low gains: unexpected contact => only small forces

28. MURI Low-Level Control Model-Based Position Control Law Plant Model Inverse Model PD Joint torque =  = arm model + K p (x p - x) + K d (v p - v) - + xdxd x xpxp x p -x More about adaptation in high-level control section

29. Typical trajectoriesPosition error vs. gain (stiffness) Without model, error is many times plotted range => Model enables good position control with low gain MURI Low-Level Control Minimum Impedance Tracking Results Commanded path = follow “wedge” at constant velocity Actual path - low k Commanded path (  Actual path, high k) Y(m)

30. Low GainHigh Gain Tradeoff between impact force and gain (stiffness) MURI Low-Level Control Minimum Impedance Control Contact Force Results Robot probes unknown environment => unexpected contact Resulting contact force:

31. Position error vs. gain Select appropriate gains for task requirements: safety, stability vs. position accuracy MURI Low-Level Control Minimum Impedance Control Performance Tradeoffs Impact force vs. gain

32. BioMimetic Robotics MURI Berkeley-Harvard Hopkins-Stanford MURI Low-Level Control Minimum Impedance Manipulation Conclusions and Future Work Developed WAM manipulation testbed: Specified & implemented arm and controller, integrated with programming environment Created Minimum Impedance Controller, demonstrated superior performance (lower forces) in unexpected contact

33. BioMimetic Robotics MURI Berkeley-Harvard Hopkins-Stanford MURI Low-Level Control Implement simple sensing (force, contact location, vision), integrate with controller to enable manipulation Research automatic learning of arm model (cf. Shadmehr) Implement impedance learning strategies (cf. Matsuoka & Howe, Shadmehr) Build SDM grippers incorporating lessons from biology, WAM testbed (Full, Shadmehr, Cutkosky) Minimum Impedance Manipulation Conclusions and Future Work

34. Fabrication MURI Low-Level Control High-Level Control What strategies are used in insect locomotion and what are their implications? Insect locomotion studies (Berkeley Bio) New measurement capabilities (Stanford) What motor control adaptation strategies do people use and how can they be applied to robots? Compliance Learning and Strategies for Unstructured Environments (Harvard & Johns Hopkins) Implications for biomimetic robots (Harvard, Stanford) Guiding questions

35. Measurement & Sensing Application of micromachined devices for small-scale biological / biomechanical force measurements 1. Adhesion force measurements of single gecko setae 2-D piezoresistive force cantilever 2. Cockroach ground reaction force measurements Custom 3-axis force sensor arrays

36. Structure of a Gecko Foot ~10 6 setae per animal Average 4.7  m in diameter 100-1000 spatulas at tip (~  0.2  m) ~20N force per ~200mm 2 pad area Adhesion by van der Waals forces?

37. 2D Piezoresistive Force sensing Special 45  ion implantation to embed piezoresistors on surfaces and side walls.

38. Experiment & Results 1. Pressed down at tip 2. Pulled away laterally Current Progress: Interpretation of data Comparison with expected values. Typical Force Curves SEM image

39. Insect Measurement Requirements Camponotus Pennsylvanicus 1cm Drosophila Melanogaster 1mm Sensor Performance Insect Blaberus Discoidalis

40. Existing Sensor Design 64x64 sensor element array, 2x2cm On-chip CMOS signal conditioning, amplification, and multiplexing Linear dynamic range 0-1.0mN Sensitivity –In-Plane: 32V/N –Normal: 171V/N Minimum resolvable load (BW=500Hz) –In-Plane: 3.5  N –Normal: 1  N

41. Sensor Elements Wire Bond Pads Wafer may be diced into strips by cutting along dashed lines Substitute Sensor Array Installation

42. Sensor Element Design Space 100 200 300 400 500 600 700 800 900 1000 0 0 50 100 150 Flexure Thickness (  m) Flexure Length (  m) Gap  0.5mm Fail Limit = wt*10*FS = 0.6N In-Plane Sensitivity  0.1V/N Normal Sensitivity  1V/N Power Dissipation  10mW Other Design Parameters: Flexure Width, w = 100  m Shuttle Plate Width, a p = 5mm Shuttle Plate Thickness, t p = 0.5mm Piezo/Flexure Fraction,  = 0.35 Bridge Excitation, V cc = 15V Implant Dose, Q = 2 x 10 13 Ions/cm 3 Min. Feature Size, m = 15  m

43. Why these measurements are important Improve S/N and add multi-axis capability. Insert MEMS approaches into Locomotion Studies, and mix Biologists and Engineers Enable progression towards smaller animals, such as ants and fruit flies.

44. Inserting sensors into SDM-manufactured limbs There are many sensors distributed throughout roach limbs, although their use in roach locomotion is not clear. SDM enables insertion of “sensing objects”, such as thermometers, strain gauges, and contact sensors. The signals from these sensors must be multiplexed and digitized, and might be reduced to “single-bit” outputs by comparing with thresholds

45. Sensor modules can be built in the form of flexible circuit hybrids, and added to the structure in the middle of SDM Inserting sensors into SDM-manufactured limbs