01 / The flywheel
Every failure in the field makes the next policy better.
This is the loop I build toward, with one of my open-source tools at each step. A deployed robot slips. ROS Time Machine and a VLM catch it. Forge and Anvil turn it into clean training data. Armature tests the new candidate. Anvil and an agent with Robotics Agent Skills ship it back to the fleet.
fleet success 72% · policy v1.0
Robot learning flywheel
An animated loop of six stations. Deploy: a robot arm on the fleet drops a cup. Capture: ROS Time Machine and a VLM save the clip around the failure. Curate: Forge scores and dedupes episodes and Anvil rejects a drifted session. Train: fine-tune on base plus field data. Evaluate: Armature runs seeded sim trials. Ship: Anvil checks the cell and an AI agent with robotics skills generates ROS 2 bringup, then the policy ships and fleet success rises.
fleet · deployment success
72%
policy v1.0 in production
illustrative numbers
01 · DEPLOY
fleet · ROS 2 · v1.0
success
72%
bin
✗ grasp slip
02 · CAPTURE
ros-time-machine + VLM
VLM · 1 fps
“cup slips from gripper” · 0.93
15 s ring buffer · RAM
now
T−5s
T+3s
T
clip_0412.mcap · event.json
03 · CURATE
forge · anvil
episode
forge score
anvil
→ train set +2 clean · −1 drift · −1 dup
04 · TRAIN
base + field data
loss
epoch 12/12
candidate v1.1
05 · EVALUATE
armature · sim
stress trials
46/50
current: 41/50
✓ beats current
same task, randomized seeds, new policy
06 · SHIP
anvil · agent skills
anvil · real cell vs pin
light · camera pose · objects
✓
agent + robotics skills
ros2 bringup · lifecycle · QoS
✓
ship v1.1 → fleet