Sim-to-real transfer

[andreaprotopapa]

Optitrack observations

• [ ] actually we used 4 cameras → we should increase the total number of cameras and being sure of seeing always each marker with at least 2 camers
• [ ] #3
• [ ] consistency of the reference system, same origin of Cartesian axis between the sim system and optitrack one
  • in sim it is set to the center of the base of the trunk.
• [ ] check the frequency of recording in optitrack for consistency with the policy control
  • is it the "Column1" the timeframe? Could it be used?

Actuators calibration

• [ ] #2

• [x] check velocity cable displacement

  • Maximum Speed: 0.166s/60° (7.4v)

• [ ] check physical bounds of control frequency

  • in the arduino code we set delays (1.5 sec of delay for each action)
  • [ ] could be speed up the frequency in the real setting?

• Options for matching in simulation

  • [x] we can diminish in sim the control frequency for matching real limits (1-2 Hz)
  • [ ] parametrize the speed in sim e we estimate with rf-dropo
  • [ ] try a fine step smaller than 1mm for 5 times
  • [ ] actuation randomization (noise)
  • sensoring feedback could be useful! however, if not available we can:
  • diminish frequency in sim
  • randomize actions (displacement cable)

• cable displacement consistency

  • cableL.addObject('CableConstraint' if not self.inverseMode else 'CableActuator', template='Vec3d',name="cable", hasPullPoint="0", indices=idx, maxPositiveDisp='70', maxDispVariation="1", minForce=0)
  • maximum displacement of 70mm in simulation

Initial position

Initial position of trunk in simulation must be consistence of reality

• trunk initial position suffers of the gravity! this is the initial position actually at regime:

• [ ] fix real reset mechanism, to replay always the initial configuration at the begin of the episode, based on fixing the initial cable length
• Options for matching in simulation
  • [ ] we set the trunk at regime applying gravity
  • [ ] random initial position of the trunk: after we are at regime, we applay a set (e.g., 3) of random actions before staring the training
• [ ] decide the position of cube, if under the trunk or close to it
• [ ] goal fixed also in reality

[andreaprotopapa]

Experimental Protocol

Experimental_Protocol___Journal.pdf

[andreaprotopapa]

Mission Report: February 21st - February 23rd, Lille INRIA

Mission Report - Lille.pdf

[andreaprotopapa]

Other TODOs by DEFROST side

• [ ] initial position is changing after one experiment
  • tubing problem
  • tension of cable problem
• [ ] new design of racks to give more tension and decide if change the motor setting with max/min release
• [ ] refine legs and floor, with the correct height corresponding to the simulation
• [ ] real-time data recording for policy usage and evaluation.
  • for this point Gabriele told me that is better using ROS doing the streaming of the data on a ROS topic. Then on the controller we get from ROS the values
• [ ] put the motor bounds in Arduino to be multiple of the RANGE
• [ ] New sofa scene, to be finished and used for training in simulation