huangyanpei/VLA
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update Wall-X to 1.1.0 (#104)

Browse Source
This commit is contained in:
Starrick Liu
2026-06-15 11:40:00 +08:00
committed by GitHub
parent e23a586846
commit 72834e7de5
200 changed files with 33916 additions and 16771 deletions
Download Patch File Download Diff File Expand all files Collapse all files
workspace/example/lerobot/qwen2_5_lerobot_template.yml
+93
View File
@@ -0,0 +1,93 @@
# Public Wall-X Qwen2.5 + LeRobot training template.
#
# Replace all /path/to/... values before launching training.
model_type: qwen2_5
task:
dof_config:
master_right_ee_cartesian_pos: 3
master_right_ee_rotation: 3
master_right_gripper: 1
agent_pos_config:
follow_right_ee_cartesian_pos: 3
follow_right_ee_rotation: 3
follow_right_gripper: 2
action_horizon: 32
action_horizon_flow: 32
use_state_string_representation: false
model:
backbone: qwen2_5
config_path: workspace/models_config/qwen2_5_moe_flash.json
processor_path: Qwen/Qwen2.5-VL-3B-Instruct
pretrained_path: Qwen/Qwen2.5-VL-3B-Instruct
attn_deterministic: true
use_ema: false
flow_loss_weight: 1.0
enable_customized_robot_config: true
customized_robot_config:
name: your_robot
customized_dof_config:
action: 7
customized_agent_pos_config:
observation.state: 8
hyperparams:
num_epoch: 100
batch_size_per_gpu: 8
gradient_accumulation_steps: 2
seed: 10233
optimizer:
optimizer_type: dmuon
learning_rate: 0.0001
max_grad_norm: 1.0
enable_grad_clip: true
betas: [0.9, 0.95]
weight_decay: 1.0e-8
eps: 1.0e-8
scheduler:
scheduler_type: cosine
num_warmup_steps: 1000
num_training_steps: 64000000
min_lr: 1.0e-5
distributed:
use_fsdp: true
use_mixed_precision: true
bf16: true
data:
dataset_type: lerobot
lerobot_config:
repo_id: /path/to/your/lerobot_dataset
root: null
key_mappings:
camera:
observation.images.image: face_view
observation.images.image2: right_wrist_view
state: observation.state
action: action
norm_stats_path: /path/to/your/norm_stats.json
train_test_split: 0.95
num_workers: 4
resolution:
face_view: 256
left_wrist_view: 256
right_wrist_view: 256
logging:
log_name: wallx_lerobot_train
log_project: wallx_public
log_entity: null
use_wandb: false
log_interval: 10
save_interval: 1000000
val_interval: 1000000
epoch_save_interval: 1
checkpoint:
save_path: /path/to/output/checkpoints
debug:
profile: false
nvtx: false
workspace/example/libero.yml
+125
View File
@@ -0,0 +1,125 @@
# LIBERO single-arm finetune example (Euler delta action, 256px, 2 cameras).
#
# Replace every /path/to/* placeholder before training:
#
# model.config_path -> model architecture JSON (mot_flash_mask_causal_xloss.json)
# model.processor_path -> Qwen2.5-VL-3B-Instruct directory
# model.pretrained_path -> same as processor_path, or HuggingFace cache path
# data.lerobot_config.repo_id -> local LeRobot dataset root (libero_all)
# data.norm_stats_path -> q01/q99 normalization JSON
# checkpoint.save_path -> writable directory for training checkpoints
# checkpoint.resume_from -> Wall-OSS-0.5 pretrained .safetensors or checkpoint directory
#
# Compute norm stats first:
#
# python scripts/compute_norm_stats.py \
# --train_config workspace/example/libero.yml \
# --data_root /path/to/libero_all \
# --output_path /path/to/libero_all_norm_stats.json
#
# Launch training (from repo root):
#
# torchrun --nproc_per_node=<N> wall_x/trainer/fsdp_trainer/train_fsdp.py \
# --config workspace/example/libero.yml
#
# Strategy: keep dof / agent_pos totals at 26 to match the pretraining action space
# via ``action_padding``. The lerobot collator right-pads libero's 7-dim action /
# 8-dim state with zeros; loss does not flow through the padded tail.
model_type: qwen2_5
task:
# Libero delta action: pos3 + rot3 + gripper1 = 7, plus action_padding(19) = 26.
dof_config:
master_right_ee_cartesian_pos: 3 # delta position
master_right_ee_rotation: 3 # delta rotation (ZYX euler)
master_right_gripper: 1
action_padding: 19
ar_dof_config:
master_right_ee_cartesian_pos: 3
master_right_ee_rotation: 3
master_right_gripper: 1
action_padding: 19
# State: pos3 + rot3 + gripper2 = 8, plus action_padding(18) = 26.
agent_pos_config:
follow_right_ee_cartesian_pos: 3
follow_right_ee_rotation: 3
follow_right_gripper: 2
action_padding: 18
action_horizon: 10
action_horizon_flow: 10
use_state_string_representation: false
model:
backbone: qwen2_5
config_path: /path/to/wall-oss-0.5/config.json
processor_path: /path/to/Qwen2.5-VL-3B-Instruct
pretrained_path: /path/to/Qwen2.5-VL-3B-Instruct
attn_deterministic: true
use_ema: false
flow_loss_weight: 1.0
ar_loss_weight: 0.01
hyperparams:
num_epoch: 100
batch_size_per_gpu: 4
gradient_accumulation_steps: 4
seed: 10222
optimizer:
optimizer_type: adamw
learning_rate: 5.0e-05
max_grad_norm: 1.0
enable_grad_clip: true
betas: [0.9, 0.95]
weight_decay: 1.0e-8
eps: 1.0e-8
scheduler:
scheduler_type: cosine
num_warmup_steps: 1000
num_training_steps: 200000
min_lr: 1.0e-6
distributed:
use_fsdp: true
use_mixed_precision: true
bf16: true
data:
dataset_type: lerobot
lerobot_config:
repo_id: /path/to/libero_all
root: null
key_mappings:
# libero_all v3.0 only has faceImg + rightImg (no leftImg).
camera:
observation.images.faceImg: face_view
observation.images.rightImg: right_wrist_view
state: observation.state
action: action
norm_stats_path: /path/to/libero_all_norm_stats.json
train_test_split: 0.95
num_workers: 4
max_length: 1024
resolution:
face_view: 256
right_wrist_view: 256
logging:
log_name: libero_ft
log_project: lerobot_libero_ft
log_entity: your_wandb_entity
use_wandb: true
log_interval: 10
save_interval: 2000
val_interval: 1000000
epoch_save_interval: 1
checkpoint:
save_path: /path/to/libero
# Single-file .safetensors loads as pretrain weights before FSDP wrapping.
# Use a checkpoint directory for full resume (optimizer / scheduler / RNG).
resume_from: /path/to/wall-oss-0.5/model.safetensors
debug:
profile: false
nvtx: false
workspace/example/maniparena_example.yml
+134
View File
@@ -0,0 +1,134 @@
# ManipArena / CVPR dual-arm finetune example (6D relative action, 448px, 3 cameras).
#
# Replace every /path/to/* placeholder before training:
#
# model.config_path -> model architecture JSON (mot_flash_mask_causal_xloss.json)
# model.processor_path -> Qwen2.5-VL-3B-Instruct directory
# model.pretrained_path -> same as processor_path, or HuggingFace cache path
# data.lerobot_config.repo_id -> local LeRobot dataset root (cvpr_4tasks)
# data.norm_stats_path -> q01/q99 normalization JSON
# checkpoint.save_path -> writable directory for training checkpoints
# checkpoint.resume_from -> Wall-OSS-0.5 pretrained .safetensors or checkpoint directory
#
# Compute norm stats first:
#
# python scripts/compute_norm_stats.py \
# --train_config workspace/example/maniparena_example.yml \
# --data_root /path/to/cvpr_4tasks \
# --output_path /path/to/cvpr_4tasks_norm_stats.json
#
# Launch training (from repo root):
#
# torchrun --nproc_per_node=<N> wall_x/trainer/fsdp_trainer/train_fsdp.py \
# --config workspace/example/maniparena_example.yml
#
# Strategy: keep dof / agent_pos totals at 26 to match the pretraining action space
# via ``action_padding``. The lerobot collator right-pads real action / state dims;
# loss does not flow through the padded tail.
model_type: qwen2_5
task:
# Dual-arm 6D relative action: 10 + 10 = 20, plus action_padding(6) = 26.
dof_config:
follow_left_ee_cartesian_pos_relative: 3
follow_left_ee_rotation_6D_relative: 6
follow_left_gripper: 1
follow_right_ee_cartesian_pos_relative: 3
follow_right_ee_rotation_6D_relative: 6
follow_right_gripper: 1
action_padding: 6
ar_dof_config:
follow_left_ee_cartesian_pos_relative: 3
follow_left_ee_rotation_6D_relative: 6
follow_left_gripper: 1
follow_right_ee_cartesian_pos_relative: 3
follow_right_ee_rotation_6D_relative: 6
follow_right_gripper: 1
action_padding: 6
agent_pos_config:
follow_left_ee_cartesian_pos: 3
follow_left_ee_rotation_6D: 6
follow_left_gripper: 1
follow_right_ee_cartesian_pos: 3
follow_right_ee_rotation_6D: 6
follow_right_gripper: 1
action_padding: 6
action_horizon: 32
action_horizon_flow: 32
use_state_string_representation: false
model:
backbone: qwen2_5
config_path: /path/to/wall-oss-0.5/config.json
processor_path: /path/to/Qwen2.5-VL-3B-Instruct
pretrained_path: /path/to/Qwen2.5-VL-3B-Instruct
attn_deterministic: true
use_ema: false
flow_loss_weight: 1.0
ar_loss_weight: 0.01
hyperparams:
num_epoch: 100
batch_size_per_gpu: 4
gradient_accumulation_steps: 4
seed: 10222
optimizer:
optimizer_type: adamw
learning_rate: 5.0e-05
max_grad_norm: 1.0
enable_grad_clip: true
betas: [0.9, 0.95]
weight_decay: 1.0e-8
eps: 1.0e-8
scheduler:
scheduler_type: cosine
num_warmup_steps: 1000
num_training_steps: 200000
min_lr: 1.0e-6
distributed:
use_fsdp: true
use_mixed_precision: true
bf16: true
data:
dataset_type: lerobot
lerobot_config:
repo_id: /path/to/cvpr_4tasks
root: null
key_mappings:
camera:
observation.images.faceImg: face_view
observation.images.leftImg: left_wrist_view
observation.images.rightImg: right_wrist_view
state: observation.state
action: action
norm_stats_path: /path/to/cvpr_4tasks_norm_stats.json
train_test_split: 0.95
num_workers: 4
max_length: 1024
resolution:
face_view: 448
left_wrist_view: 448
right_wrist_view: 448
logging:
log_name: maniparena_ft
log_project: lerobot_maniparena_ft
log_entity: your_wandb_entity
use_wandb: true
log_interval: 10
save_interval: 2000
val_interval: 1000000
epoch_save_interval: 1
checkpoint:
save_path: /path/to/cvpr_4tasks
# Single-file .safetensors loads as pretrain weights before FSDP wrapping.
# Use a checkpoint directory for full resume (optimizer / scheduler / RNG).
resume_from: /path/to/wall-oss-0.5/model.safetensors
debug:
profile: false
nvtx: false