Verified80% confidenceFactExact time
Behavioral cloning suffers from the distribution shift problem: during training the policy only encounters the state distribution covered by expert trajectories, and during deployment the accumulation of small deviations causes the actual state distribution visited to gradually diverge from the training distribution.
4
Sources
80%
Confidence
Long-term
Relevance
8/17/2026
First Seen
Sources
机械手抓取任务失败？强化学习新手常见陷阱与解决方案
Related Claims
UnverifiedDistribution Shift refers to the fact that the behavior an Agent exhibits during training and controlled testing may not represent its behavior in a real deployment environment.79% similarUnverifiedIn distributed training, a small number of nodes whose kernel parameters are inconsistent with the rest can drag down the throughput of the entire training job, creating bottleneck nodes.78% similarUnverifiedAsking the demonstrator to adjust their posture may seem like a low-cost compromise, but it quietly changes the distribution of the training data.76% similarUnverified训练-部署鸿沟指模型因训练与部署环境的数据分布偏移、软件依赖版本差异或硬件指令集不同而导致的性能下降76% similarUnverifiedWhen a neural network trained offline is deployed to a closed-loop control system, its actions change the system's state distribution, causing the data distribution at actual runtime to deviate from the offline data distribution at training time, that is, distribution shift.76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/763754API
curl https://kongchang.com/api/v1/knowledge/claims/763754MCP
get_claim(id=763754)