Robotics
LeRobot
English
OpenRAL
rskill
diffusion
vision-language-action
pusht
diffusion-policy
manipulation
Instructions to use OpenRAL/rskill-diffusion-pusht-pusht-fp32 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LeRobot
How to use OpenRAL/rskill-diffusion-pusht-pusht-fp32 with LeRobot:
- Notebooks
- Google Colab
- Kaggle
rskill.yaml: sync to OpenRAL namespace + current schema (joint_units, evaluated_tasks, ADR-0071)
Browse files- rskill.yaml +19 -12
rskill.yaml
CHANGED
|
@@ -5,6 +5,7 @@
|
|
| 5 |
# ββ Identity βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 6 |
schema_version: "0.1"
|
| 7 |
name: "OpenRAL/rskill-diffusion-pusht"
|
|
|
|
| 8 |
version: "0.1.0"
|
| 9 |
license: "apache-2.0"
|
| 10 |
role: "s1"
|
|
@@ -28,17 +29,17 @@ sensors_required:
|
|
| 28 |
min_width: 96
|
| 29 |
min_height: 96
|
| 30 |
|
| 31 |
-
# Output side (ADR-0013). The pusht_2d scene-pseudo-robot exposes
|
| 32 |
-
# (x, y)
|
| 33 |
-
# `
|
| 34 |
-
#
|
|
|
|
|
|
|
| 35 |
# The loader auto-fills n_dof (2) + vla_action_key from the robot YAML.
|
| 36 |
-
# PushT actions are absolute (x, y) targets in the world/scene frame.
|
| 37 |
actuators_required:
|
| 38 |
-
- kind: "
|
| 39 |
control_mode_semantics:
|
| 40 |
mode: "absolute"
|
| 41 |
-
reference_frame: "world"
|
| 42 |
|
| 43 |
# ββ Runtime / weights ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 44 |
runtime: "pytorch"
|
|
@@ -70,7 +71,7 @@ latency_budget:
|
|
| 70 |
# ββ Provenance βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 71 |
# Headline success rate from skills/diffusion-pusht/eval/pusht.json.
|
| 72 |
benchmarks:
|
| 73 |
-
pusht:
|
| 74 |
|
| 75 |
paper_url: "https://arxiv.org/abs/2303.04137"
|
| 76 |
source_repo: "hf://lerobot/diffusion_pusht"
|
|
@@ -96,8 +97,14 @@ scenes:
|
|
| 96 |
# to bind the LeRobot v3 `action` feature shape).
|
| 97 |
action_contract:
|
| 98 |
dim: 2
|
| 99 |
-
#
|
| 100 |
-
#
|
| 101 |
-
#
|
| 102 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 103 |
joint_units: "radians"
|
|
|
|
| 5 |
# ββ Identity βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 6 |
schema_version: "0.1"
|
| 7 |
name: "OpenRAL/rskill-diffusion-pusht"
|
| 8 |
+
evaluated_tasks: ["pusht"] # ADR-0060: benchmark task this checkpoint is validated for (gate)
|
| 9 |
version: "0.1.0"
|
| 10 |
license: "apache-2.0"
|
| 11 |
role: "s1"
|
|
|
|
| 29 |
min_width: 96
|
| 30 |
min_height: 96
|
| 31 |
|
| 32 |
+
# Output side (ADR-0013). The pusht_2d scene-pseudo-robot exposes its 2-D
|
| 33 |
+
# (x, y) pusher position as two synthetic prismatic joints (tip_x, tip_y);
|
| 34 |
+
# robots/pusht_2d/robot.yaml advertises `joint_position` over those 2 DoF.
|
| 35 |
+
# The gym-pusht action IS the absolute target of those joints, so the
|
| 36 |
+
# robot-side actuator is joint_position (ADR-0071 β matches every other
|
| 37 |
+
# rSkill, where `kind` names the robot actuator, not the policy encoding).
|
| 38 |
# The loader auto-fills n_dof (2) + vla_action_key from the robot YAML.
|
|
|
|
| 39 |
actuators_required:
|
| 40 |
+
- kind: "joint_position"
|
| 41 |
control_mode_semantics:
|
| 42 |
mode: "absolute"
|
|
|
|
| 43 |
|
| 44 |
# ββ Runtime / weights ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 45 |
runtime: "pytorch"
|
|
|
|
| 71 |
# ββ Provenance βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 72 |
# Headline success rate from skills/diffusion-pusht/eval/pusht.json.
|
| 73 |
benchmarks:
|
| 74 |
+
pusht: 1.0
|
| 75 |
|
| 76 |
paper_url: "https://arxiv.org/abs/2303.04137"
|
| 77 |
source_repo: "hf://lerobot/diffusion_pusht"
|
|
|
|
| 97 |
# to bind the LeRobot v3 `action` feature shape).
|
| 98 |
action_contract:
|
| 99 |
dim: 2
|
| 100 |
+
# ADR-0071 β absolute positions of the 2 prismatic tip_x/tip_y joints.
|
| 101 |
+
# Declared explicitly so the contract states joint-space rather than relying
|
| 102 |
+
# on the undeclared-layout fallback.
|
| 103 |
+
representation: "joint_positions"
|
| 104 |
+
# EXPLICIT joint units β PushT's 2-D action is a planar tip position in the
|
| 105 |
+
# 512Γ512 image frame (verified: observation.state normalizer maxβ[496, 511],
|
| 106 |
+
# minβ[13, 33]), NOT angles. "radians" selects the identity (no degβrad
|
| 107 |
+
# conversion) so the 0β512 values pass through unchanged. It is declared to
|
| 108 |
+
# STOP the stats-magnitude heuristic β which sees peak β«Ο and would wrongly
|
| 109 |
+
# flag "degrees" and corrupt the action with a ΓΟ/180 scaling.
|
| 110 |
joint_units: "radians"
|