사족보행 로봇 연구 기록
← 요약으로 돌아가기
V1 · 소프트웨어 · Oct 12, 2026

Multi-terrain training: three 1 B-step runs

Three runs on the multi-terrain field, 1 billion steps each, measured on the test course beside our stair policy. None clears the bar of 5 points above it on the trained tests, so the stair policy stays.

  • Trained / held-out tests: stair policy 83.6 % / 85.8 %, run C (from the stair policy) 82.9 % / 84.9 %, run B (from the flat policy) 78.8 % / 81.3 %, run A (flat start, plain critic) 75.4 % / 71.4 %
  • Training held the obstacles at 2.4–2.9 cm stairs and ±2.1–2.5 cm boxes; the hard test lanes are 4–6 cm
  • The critic's simulator inputs helped most on the hardest tests: 4 cm stairs 68 vs 11 of 100 (B vs A)
  • 5 cm stairs: 36 of 100 for run C, 17 for the stair policy