Tag: #강화학습

← All posts

• Reinforcement Learning Evaluation: Training Balls and Evaluation Balls, Medians Across Multiple Seeds, Final Value Reporting, and Baselines

  Reinforcement learning results are easily inflated if you only look at a single seed or the single best moment. We re-examined 48 training runs across posts 11 to 14 to review the evaluation rules with numbers. Under the

  2026-10-18 · 10 min read #탁구 로봇#강화학습#평가#재현성#시드

• From Simulation to Reality: Domain Randomization, Delay and Noise Modeling, System Identification, and Massively Parallel Simulation

  Policies learned in a simulator fail in the real world to the extent that the simulator is wrong. For the pendulum racket problem, we created a separate "reality" with a racket restitution of 0.72, observation delay of 2

  2026-10-18 · 11 min read #탁구 로봇#강화학습#sim2real#도메인 랜덤화#시스템 식별

• Reward Design and Action Space: Sparse vs. Shaped Rewards, Reward Hacking, Joint Limit Penalties, Direct Control vs. Hybrid

  Reinforcement learning only learns what the reward tells it. In the pendulum racket experiment of Episode 12, five different rewards were tested using PPO. A sparse reward of "1 for scoring" achieved 87%, while a shaped

  2026-10-18 · 12 min read #탁구 로봇#강화학습#보상 설계#보상 해킹#행동 공간

• Policy Gradients and PPO: Why We Clip Ratios, Advantage Functions and GAE, Value Functions, Entropy

  Policy gradients increase the probability of actions that performed better than usual, and PPO reuses the same data multiple times by clipping the probability ratio between the new and old policies so they do not move to

  2026-10-18 · 12 min read #탁구 로봇#강화학습#PPO#정책 경사#GAE

• Classical Methods and Reinforcement Learning: What to Solve with Equations and What to Leave to Learning

  All table tennis robot papers mix equations and learning, but draw the boundary at different places. Mülling et al. 2013 left action selection to learning, Büchler et al. 2022 left everything down to muscle control, and

  2026-10-18 · 15 min read #탁구 로봇#강화학습#하이브리드 제어#시스템 식별#정책 탐색

• The Framework of Reinforcement Learning: MDP, State, Action, Reward, Discount, Policy, and Value Function Read Through a Table Tennis Arm

  Reinforcement learning repeats the process of observing states, choosing actions, and receiving rewards to find a policy that maximizes the sum of rewards. This post maps the five elements of that framework (MDP) one by

  2026-10-18 · 12 min read #탁구 로봇#강화학습#MDP#가치 함수#할인율

• Until a Table Tennis Robot Returns a Ball: Calculations Seen Through a Step-by-Step Simulation Video

  Recorded using MuJoCo, the process from when the table tennis robot arm receives an opponent's shot and places it at the target position on the opponent's court is illustrated across seven stages: observation, prediction

  2026-10-18 · 31 min read #탁구 로봇#MuJoCo#시뮬레이션 영상#역기구학#궤적 생성