V1 · 프로젝트 정보
Building Pupper v3, training locomotion policies and bringing them to the robot.
프로젝트 현황
Printed parts
Done
28 parts; ears and strap left out
Stairs crossed (simulation)
3 cm
97 of 100 crossings on average; 4 cm: 79
Standing height (simulation)
0.176 m
official policy 0.183 m
1 B training steps
46 min
lab PC, MuJoCo Warp
다음 작업
하드웨어 · 다음 작업
Finishing the assembly, then bring-up; the stair course is designed and waits for its rise to be picked (30 or 40 mm)
소프트웨어 · 다음 작업
Training on grass; behaviour commands (training with vertical stair edges and with gear backlash, the crouch on slopes); one policy for walking, commands and fall recovery (its stand-up after a fall in the robot's control software); the heading hold on the robot (IMU heading near the motors); then the deployment check
로드맵
하드웨어
- Base designAug 31 · Pupper v3 selectedDone
- 3D printingOct 5 · 28 parts; ears and strap left outDone
- Parts arrivalby Oct 22 · motors, boards, Raspberry Pi 5Done
- Assemblyrecorded from above on Oct 22In progress
- Bring-upfirmware, zeroing, IMUNot started
- First walkshipped policyNot started
- Terrain printedstair course designed; same CAD as the simulatorNot started
소프트웨어
- Policy trained in simulationOct 2 · 1B stepsDone
- ROS 2 stack analysedOct 2 · robot control flowDone
- Policy in the ROS 2 stackOct 2 · simulation, keys or padDone
- Match the official policyOct 12 · height 0.176 m, no fall in 10 runsDone
- Terrain training3 cm stairs crossed in simulation; multi-terrain and behaviour runs openIn progress
- Deployment checkexport, e-stop, Pi 5 timingNot started
- Motor benchafter parts arriveNot started
공통 목표
- Our policy on the robotflat groundNot started
- Our policy on the printed terrainsame layout as the simulatorNot started
- V1.0 demo: spoken commandssit, stand, follow me, lie down; plan of Oct 28Not started
자료
- Pupper v3 documentation
- pupperv3-monorepo
- pupperv3-mjx
- World-model data (robot recordings with actions
- Robot camera recordings (YOLO detections)
- Report: scaling RL policies versus BC
- Report: autonomous-intelligence architecture and Shappy
- Report: gaits of dogs and quadruped robots
- Report: momentum beyond the torque limit
- Report: perceptive foothold selection
- Report: aerial reorientation and safe landing
- Report: multi-terrain policy training
- Multi-terrain training
- Stair environment
- ROS 2 graph in simulation