Training Update

[tuananhroman]

• more robust model weight transfer
• adjusted training parameters
• fix robot radius namespacing
• square root learning rate decay
• fix semantic info namespacing during evaluation
• more accurate vectorized stats recorder
• update dockerfile