Robotics
LeRobot
English
OpenRAL
rskill
smolvla
vision-language-action
aloha_agilex
vla
robotwin
bimanual
manipulation
Instructions to use OpenRAL/rskill-smolvla-aloha_agilex-robotwin-bf16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LeRobot
How to use OpenRAL/rskill-smolvla-aloha_agilex-robotwin-bf16 with LeRobot:
# See https://github.com/huggingface/lerobot?tab=readme-ov-file#installation for more details git clone https://github.com/huggingface/lerobot.git cd lerobot pip install -e .[smolvla]
# Launch finetuning on your dataset python lerobot/scripts/train.py \ --policy.path=OpenRAL/rskill-smolvla-aloha_agilex-robotwin-bf16 \ --dataset.repo_id=lerobot/svla_so101_pickplace \ --batch_size=64 \ --steps=20000 \ --output_dir=outputs/train/my_smolvla \ --job_name=my_smolvla_training \ --policy.device=cuda \ --wandb.enable=true
# Run the policy using the record function python -m lerobot.record \ --robot.type=so101_follower \ --robot.port=/dev/ttyACM0 \ # <- Use your port --robot.id=my_blue_follower_arm \ # <- Use your robot id --robot.cameras="{ front: {type: opencv, index_or_path: 8, width: 640, height: 480, fps: 30}}" \ # <- Use your cameras --dataset.single_task="Grasp a lego block and put it in the bin." \ # <- Use the same task description you used in your dataset recording --dataset.repo_id=HF_USER/dataset_name \ # <- This will be the dataset name on HF Hub --dataset.episode_time_s=50 \ --dataset.num_episodes=10 \ --policy.path=OpenRAL/rskill-smolvla-aloha_agilex-robotwin-bf16 - Notebooks
- Google Colab
- Kaggle
chore: canonical naming migration
Browse files- rskill.yaml +10 -10
rskill.yaml
CHANGED
|
@@ -25,7 +25,7 @@
|
|
| 25 |
# RoboTwin's native LeRobot env cameras are head_camera / left_camera /
|
| 26 |
# right_camera. The sidecar (tools/robotwin_sidecar.py) re-keys them, in fixed
|
| 27 |
# headβleftβright order, to the SCENE's camera names β the aloha_agilex canonical
|
| 28 |
-
# sensor names top / wrist_left / wrist_right
|
| 29 |
# checkpoint expects observation.images.camera1/2/3, so the sim adapter needs a
|
| 30 |
# rename from the scene names to the checkpoint keys. That rename is declared per
|
| 31 |
# skill in `image_preprocessing.aliases` below β the same mechanism smolvla-libero
|
|
@@ -42,7 +42,7 @@
|
|
| 42 |
|
| 43 |
# ββ Identity βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 44 |
schema_version: "0.1"
|
| 45 |
-
name: "OpenRAL/rskill-smolvla-robotwin"
|
| 46 |
version: "0.1.0"
|
| 47 |
license: "apache-2.0"
|
| 48 |
role: "s1"
|
|
@@ -73,7 +73,7 @@ sensors_required:
|
|
| 73 |
min_width: 256
|
| 74 |
min_height: 256
|
| 75 |
|
| 76 |
-
# Output side
|
| 77 |
# the loader auto-fills n_dof (14) + vla_action_key from
|
| 78 |
# robots/aloha_agilex/robot.yaml.
|
| 79 |
actuators_required:
|
|
@@ -86,7 +86,7 @@ runtime: "pytorch"
|
|
| 86 |
quantization:
|
| 87 |
dtype: "bf16"
|
| 88 |
backend: "pytorch"
|
| 89 |
-
weights_uri: "hf://lerobot/smolvla_robotwin
|
| 90 |
|
| 91 |
# ββ Preprocessing (all knobs needed to interpret IO) βββββββββββββββββββββββ
|
| 92 |
# The official checkpoint ships the lerobot PolicyProcessorPipeline JSONs
|
|
@@ -125,7 +125,7 @@ latency_budget:
|
|
| 125 |
# ββ Provenance βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 126 |
# No benchmarks shipped yet β see eval/.gitkeep. Populated by
|
| 127 |
# `openral benchmark run --suite robotwin` once the SAPIEN sidecar venv is
|
| 128 |
-
# provisioned
|
| 129 |
# reference (Pi0 46.4/16.3 Easy/Hard) are paper-cited, not reproduced here.
|
| 130 |
paper_url: "https://arxiv.org/abs/2506.18088"
|
| 131 |
dataset_uri: "hf://lerobot/robotwin_unified"
|
|
@@ -136,16 +136,16 @@ description: >
|
|
| 136 |
unified dataset (50 dual-arm tasks, aloha-agilex embodiment, SAPIEN).
|
| 137 |
Multi-task: action chunks of length 50 across three RGB views
|
| 138 |
(head + per-wrist) driving a 14-DoF dual-arm joint command. Runs on the
|
| 139 |
-
RoboTwin scene backend through the out-of-process SAPIEN sidecar
|
| 140 |
|
| 141 |
-
# ββ Task-data gate
|
| 142 |
# Multi-task checkpoint trained on the full robotwin_unified set, so it is
|
| 143 |
# valid on every RoboTwin scene. Declaring the scene id `robotwin` makes the
|
| 144 |
# benchmark runner accept all `robotwin/<task>` scenes (family match) and
|
| 145 |
# refuse non-RoboTwin scenes.
|
| 146 |
evaluated_tasks: ["robotwin"]
|
| 147 |
|
| 148 |
-
# ββ Action vocabulary
|
| 149 |
# GENERALIST marks the multi-task foundation checkpoint; the specific verbs
|
| 150 |
# cover the bimanual-coordination / pick-place / stacking task families.
|
| 151 |
actions:
|
|
@@ -161,10 +161,10 @@ objects:
|
|
| 161 |
scenes:
|
| 162 |
- "tabletop"
|
| 163 |
|
| 164 |
-
# ββ Per-checkpoint action contract
|
| 165 |
action_contract:
|
| 166 |
dim: 14
|
| 167 |
-
#
|
| 168 |
# (2Γ(6 arm + 1 gripper)).
|
| 169 |
representation: "joint_positions"
|
| 170 |
# EXPLICIT joint units β RoboTwin (SAPIEN, aloha-agilex qpos) records in
|
|
|
|
| 25 |
# RoboTwin's native LeRobot env cameras are head_camera / left_camera /
|
| 26 |
# right_camera. The sidecar (tools/robotwin_sidecar.py) re-keys them, in fixed
|
| 27 |
# headβleftβright order, to the SCENE's camera names β the aloha_agilex canonical
|
| 28 |
+
# sensor names top / wrist_left / wrist_right, NOT camera1/2/3. The
|
| 29 |
# checkpoint expects observation.images.camera1/2/3, so the sim adapter needs a
|
| 30 |
# rename from the scene names to the checkpoint keys. That rename is declared per
|
| 31 |
# skill in `image_preprocessing.aliases` below β the same mechanism smolvla-libero
|
|
|
|
| 42 |
|
| 43 |
# ββ Identity βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 44 |
schema_version: "0.1"
|
| 45 |
+
name: "OpenRAL/rskill-smolvla-aloha_agilex-robotwin-bf16"
|
| 46 |
version: "0.1.0"
|
| 47 |
license: "apache-2.0"
|
| 48 |
role: "s1"
|
|
|
|
| 73 |
min_width: 256
|
| 74 |
min_height: 256
|
| 75 |
|
| 76 |
+
# Output side. 14-DoF dual-arm joint command. For aloha_agilex
|
| 77 |
# the loader auto-fills n_dof (14) + vla_action_key from
|
| 78 |
# robots/aloha_agilex/robot.yaml.
|
| 79 |
actuators_required:
|
|
|
|
| 86 |
quantization:
|
| 87 |
dtype: "bf16"
|
| 88 |
backend: "pytorch"
|
| 89 |
+
weights_uri: "hf://lerobot/smolvla_robotwin"
|
| 90 |
|
| 91 |
# ββ Preprocessing (all knobs needed to interpret IO) βββββββββββββββββββββββ
|
| 92 |
# The official checkpoint ships the lerobot PolicyProcessorPipeline JSONs
|
|
|
|
| 125 |
# ββ Provenance βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 126 |
# No benchmarks shipped yet β see eval/.gitkeep. Populated by
|
| 127 |
# `openral benchmark run --suite robotwin` once the SAPIEN sidecar venv is
|
| 128 |
+
# provisioned. RoboTwin leaderboard SmolVLA-class baselines for
|
| 129 |
# reference (Pi0 46.4/16.3 Easy/Hard) are paper-cited, not reproduced here.
|
| 130 |
paper_url: "https://arxiv.org/abs/2506.18088"
|
| 131 |
dataset_uri: "hf://lerobot/robotwin_unified"
|
|
|
|
| 136 |
unified dataset (50 dual-arm tasks, aloha-agilex embodiment, SAPIEN).
|
| 137 |
Multi-task: action chunks of length 50 across three RGB views
|
| 138 |
(head + per-wrist) driving a 14-DoF dual-arm joint command. Runs on the
|
| 139 |
+
RoboTwin scene backend through the out-of-process SAPIEN sidecar.
|
| 140 |
|
| 141 |
+
# ββ Task-data gate βββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 142 |
# Multi-task checkpoint trained on the full robotwin_unified set, so it is
|
| 143 |
# valid on every RoboTwin scene. Declaring the scene id `robotwin` makes the
|
| 144 |
# benchmark runner accept all `robotwin/<task>` scenes (family match) and
|
| 145 |
# refuse non-RoboTwin scenes.
|
| 146 |
evaluated_tasks: ["robotwin"]
|
| 147 |
|
| 148 |
+
# ββ Action vocabulary ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 149 |
# GENERALIST marks the multi-task foundation checkpoint; the specific verbs
|
| 150 |
# cover the bimanual-coordination / pick-place / stacking task families.
|
| 151 |
actions:
|
|
|
|
| 161 |
scenes:
|
| 162 |
- "tabletop"
|
| 163 |
|
| 164 |
+
# ββ Per-checkpoint action contract βββββββββββββββββββββββββββββββββββββββββ
|
| 165 |
action_contract:
|
| 166 |
dim: 14
|
| 167 |
+
# RoboTwin aloha-agilex emits absolute joint positions
|
| 168 |
# (2Γ(6 arm + 1 gripper)).
|
| 169 |
representation: "joint_positions"
|
| 170 |
# EXPLICIT joint units β RoboTwin (SAPIEN, aloha-agilex qpos) records in
|