rskill.yaml

# A1 Runtime LingBot-VA checkpoint deployed through OpenRAL.
#
# The model predicts episode-relative EEF pose + normalized gripper chunks.
# The Runtime-owned policy gateway validates those targets and lowers each step
# to six absolute A1 joint positions plus one gripper value.
# Both typed actions then traverse OpenRAL's safety kernel and A1 HAL.
schema_version: "0.1"
name: "OpenRAL/rskill-lingbot_va_a1-galaxea_a1-fruit_placement-bf16"
version: "0.1.0"
license: "apache-2.0"
role: "s1"
kind: "vla"
model_family: "lingbot_va_a1"

embodiment_tags:
  - "galaxea_a1"

sensors_required:
  - modality: "rgb"
    vla_feature_key: "observation.images.front"
    min_width: 480
    min_height: 480
  - modality: "rgb"
    vla_feature_key: "observation.images.wrist"
    min_width: 640
    min_height: 480

actuators_required:
  - kind: "joint_position"
    control_mode_semantics:
      mode: "absolute"
      joint_order:
        - arm_joint1
        - arm_joint2
        - arm_joint3
        - arm_joint4
        - arm_joint5
        - arm_joint6
  - kind: "gripper_position"
    control_mode_semantics:
      mode: "absolute"
      gripper_convention: "normalized_open_unit"

runtime: "pytorch"
quantization:
  dtype: "bf16"
  backend: "pytorch"
weights_uri: "hf://pengyue-polaron/lingbot-va-galaxea-a1-fruit-placement-eef@90e017bdbc6afac2e441b4634c9192776bbcb8b7"

# No `processors` block: this checkpoint is not a lerobot
# PolicyProcessorPipeline. Quantile normalization and the action-channel map
# ship in configs/va_a1_cfg.py and are applied by the external LingBot server
# before OpenRAL receives the physical EEF target.
state_contract:
  dim: 6

chunk_size: 16
n_action_steps: 8
latency_budget:
  # Measured through the full OpenRAL camera + websocket + EEF/IK adapter:
  # ~4.05 s for the first model call on this A1 host; replay ticks are cheap.
  per_chunk_ms: 6000.0
  max_execution_s: 420.0

dataset_uri: "hf://pengyue-polaron/nyush-galaxea-a1-fruit-placement-eef-v21@1bc2c4035e7dc638f7dd9fa5ec7987bec66d0933"
source_repo: "hf://robbyant/lingbot-va-base"
description: >
  LingBot-VA fruit-placement policy for the Galaxea A1. It consumes the
  synchronized front and wrist RGB views, predicts episode-relative EEF pose
  and continuous gripper chunks, and uses the tracked A1 Runtime IK contract
  before OpenRAL validates and executes the resulting joint/gripper actions.

# The LingBot EEF solution may be farther from current feedback than the A1
# joint tracker accepts in one update. This policy-owned per-tick lowering
# bound remains below both the HAL's 0.08 rad live target-step ceiling and its
# 0.05 rad initial-alignment ceiling.
policy_extras:
  max_joint_substep_rad: 0.045

actions:
  - "pick"
  - "place"
objects:
  - "fruit"
  - "mango"
  - "bowl"
  - "plate"
scenes:
  - "tabletop"

action_contract:
  dim: 7
  joint_units: "radians"
  slots:
    - range: [0, 5]
      control_mode: "joint_position"
      joint_names:
        - arm_joint1
        - arm_joint2
        - arm_joint3
        - arm_joint4
        - arm_joint5
        - arm_joint6
    - range: [6, 6]
      control_mode: "gripper_position"
      ee: "gripper"