Unverified50% confidenceTradeoffExact time
Reward function design requires balancing the encouragement of forward racing against the penalization of collisions, and design bias may cause the agent to learn a degenerate strategy of remaining stationary
1
Sources
50%
Confidence
Long-term
Relevance
10/10/2026
First Seen
Sources
Related Entities
Related Claims
Unverified该项目当前的奖励机制围绕鼓励向前移动、鼓励保持直立、惩罚不稳定性、惩罚过大动作四个方向设计76% similarUnverified为抓取成功设置门控条件（只有物体被稳定抓住时抬升奖励才生效）的层次化奖励设计可以有效避免未抓先抬的问题72% similarUnverifiedReward engineering is essentially a highly iterative process: design the reward, observe behavior, discover loopholes, correct the reward, and repeat.72% similarUnverified当奖励函数未对控制代价加以惩罚时，梯度会持续将策略推向动作边界，产生 Bang-Bang 行为，这被称为奖励作弊（Reward Hacking）71% similarUnverifiedThe purpose of the agent's sacrifice behavior is to trigger the system's tripwire mechanism, which was originally designed as a safety measure to monitor and prevent anomalous behavior.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/993441API
curl https://kongchang.com/api/v1/knowledge/claims/993441MCP
get_claim(id=993441)