返回Glossary
梯度裁剪
在训练期间限制梯度的幅度，以防止爆炸梯度导致优化不稳定。梯度范数被裁剪到最大值（通常为 1.0–10.0）。梯度裁剪是训练 Transformer、RNN 和强化学习策略的标准做法，其中奖励方差可能导致大的梯度尖峰。
Open 7-DoF arm, bimanual-ready. Ships from San Francisco with pilot data-collection packages from $2,500.
90+ platforms — humanoids, arms, quadrupeds, dexterous hands, and teleoperation kits. In-stock hardware ships in 48 hours.
Hardware-synced demonstrations for VLA and imitation learning. Download in LeRobot, HDF5, or RLDS format.
From first arm on the bench to a policy that survives a real workcell — the Robotics Center pipeline in three steps.
Low-latency data-collection glove for dexterous manipulation. From $5,500, ships from San Francisco.
Unbox, calibrate, and run your first autonomous walk. Includes ROS 2 bringup and teleop quickstart.
在训练期间限制梯度的幅度，以防止爆炸梯度导致优化不稳定。梯度范数被裁剪到最大值（通常为 1.0–10.0）。梯度裁剪是训练 Transformer、RNN 和强化学习策略的标准做法，其中奖励方差可能导致大的梯度尖峰。