ur5-follow / config.yaml
gpue's picture
Upload folder using huggingface_hub
f31c630 verified
# Nova-Sim Observation Configuration
#
# This file defines which parameters from the Nova-Sim state stream
# are mapped to robot observations for recording, replay, and inference.
#
# To disable a feature, comment out the entire section or set enabled: false
repositories:
# Dataset repo used for upload/download and dataset metadata.
dataset: "${HF_USER}/ur5-follow"
dataset_branch: "main"
# Policy repo used for model upload/download.
policy: "${HF_USER}/ur5-follow"
paths:
data_dir: "./data"
checkpoint_dir: "./policies"
ui:
trainer_host: "127.0.0.1"
trainer_port: 3005
recording:
task_name: "follow"
episode_time_s: 5
reset_time_s: 0
num_episodes: 10
flow: "follow"
follow:
target_object: "t_object"
max_velocity: 0.5
kp: 1.0
deadzone_m: 0.01
# Action horizon: how many timesteps ahead for target position actions
# Time horizon = n_action_steps / fps (e.g., 5 steps @ 10 FPS = 0.5s)
n_action_steps: 20
robot:
type: "nova_sim"
id: "ur5_robot"
# Raw data capture (recording schema)
data:
fps: 20
fields:
# End-effector Cartesian position (in meters)
# Source: end_effector dict from Nova-Sim state
- name: ee_x
source: end_effector
field: x
- name: ee_y
source: end_effector
field: y
- name: ee_z
source: end_effector
field: z
# End-effector orientation (quaternion)
# Source: ee_orientation dict from Nova-Sim state
- name: ee_qw
source: ee_orientation
field: w
- name: ee_qx
source: ee_orientation
field: x
- name: ee_qy
source: ee_orientation
field: y
- name: ee_qz
source: ee_orientation
field: z
# End-effector target position (meters)
# Source: ee_target dict from Nova-Sim state
- name: target_x
source: ee_target
field: x
- name: target_y
source: ee_target
field: y
- name: target_z
source: ee_target
field: z
# End-effector target orientation (Euler angles)
# Source: ee_target_orientation dict from Nova-Sim state
- name: target_roll
source: ee_target_orientation
field: roll
- name: target_pitch
source: ee_target_orientation
field: pitch
- name: target_yaw
source: ee_target_orientation
field: yaw
# Target object pose from scene stream (for follow supervision/debugging)
- name: t_obj_x
source: scene_object:t_object
field: position.x
- name: t_obj_y
source: scene_object:t_object
field: position.y
- name: t_obj_z
source: scene_object:t_object
field: position.z
- name: t_obj_qw
source: scene_object:t_object
field: orientation.w
- name: t_obj_qx
source: scene_object:t_object
field: orientation.x
- name: t_obj_qy
source: scene_object:t_object
field: orientation.y
- name: t_obj_qz
source: scene_object:t_object
field: orientation.z
# Gripper state (0-255)
# Source: gripper scalar from Nova-Sim state
- name: gripper
source: gripper
# Reward (per-step)
# Source: reward scalar from Nova-Sim state
- name: reward
source: reward
# Joint positions (6)
# Source: joint_positions array from Nova-Sim state
- name: joint_0
source: joint_positions
index: 0
- name: joint_1
source: joint_positions
index: 1
- name: joint_2
source: joint_positions
index: 2
- name: joint_3
source: joint_positions
index: 3
- name: joint_4
source: joint_positions
index: 4
- name: joint_5
source: joint_positions
index: 5
# Joint targets (6)
# Source: joint_targets array from Nova-Sim state
- name: joint_target_0
source: joint_targets
index: 0
- name: joint_target_1
source: joint_targets
index: 1
- name: joint_target_2
source: joint_targets
index: 2
- name: joint_target_3
source: joint_targets
index: 3
- name: joint_target_4
source: joint_targets
index: 4
- name: joint_target_5
source: joint_targets
index: 5
# Camera streams recorded at dataset root (<camera_name>)
cameras:
fps: 20
# WACT wrist camera: attached to end-effector, straight -x view (no diagonal orbit)
ee_wrist:
width: 240
height: 160
label: WACT XY Wrist -X
follow_site: ee_site
look_target_site: stick_tip
# Straight -x with elevated orbit; use elevation/distance instead of offset.z.
azimuth: 180
elevation: -30
distance: 1
track_orientation: false
offset: [0.35, 0.0, 0.0]
preprocessing:
wact_xy:
reward:
success_reward_min: -0.05
fps: 10
observations:
fps: 10
# Keep Cartesian state channels for debugging/aux tasks.
state:
- ee_x
- ee_y
- target_x
- target_y
# Main vision stream for WACT.
images:
- ee_wrist
actions:
fps: 10
items:
- vx
- vy
training:
wact_xy:
policy_type: "act"
training_steps: 20000
policy_private: false
hf_hardware: "l4x1"
hf_timeout: "12h"
inference:
# Pose restrictions applied during replay/inference when sending actions.
policy: wact_xy
pose_restrictions:
cartesian:
enabled: true
x:
min: 0.3
max: 0.9
y:
min: -0.6
max: 0.6
z:
min: 0.7
max: 1.0
# Episode metadata (always included, cannot be disabled)
metadata:
- reward
- terminated
- truncated
- steps