39 lines
1.2 KiB
YAML
39 lines
1.2 KiB
YAML
# @package agent
|
||
defaults:
|
||
# - /backbone@vision_backbone: resnet
|
||
- /backbone@vision_backbone: resnet_diffusion
|
||
- /modules@state_encoder: identity_state_encoder
|
||
- /modules@action_encoder: identity_action_encoder
|
||
- /head: conditional_unet1d
|
||
- _self_
|
||
|
||
_target_: roboimi.vla.agent.VLAAgent
|
||
|
||
# ====================
|
||
# 模型维度配置
|
||
# ====================
|
||
action_dim: 16 # 动作维度（机器人关节数）
|
||
obs_dim: 16 # 本体感知维度（关节位置）
|
||
|
||
# ====================
|
||
#
|
||
# ====================
|
||
normalization_type: "min_max" # "min_max" or "gaussian"
|
||
|
||
# ====================
|
||
# 时间步配置
|
||
# ====================
|
||
pred_horizon: 8 # 预测未来多少步动作
|
||
obs_horizon: 2 # 使用多少步历史观测
|
||
num_action_steps: 4 # 每次推理实际执行多少步动作（应 <= pred_horizon - obs_horizon + 1）
|
||
|
||
# ====================
|
||
# 相机配置
|
||
# ====================
|
||
num_cams: 3 # 摄像头数量 (r_vis, top, front)
|
||
|
||
# ====================
|
||
# 扩散过程配置
|
||
# ====================
|
||
diffusion_steps: 100 # 扩散训练步数（DDPM）
|
||
inference_steps: 10 # 推理时的去噪步数（DDIM，固定为 10） |