Agentic 3D
Product references and dimensions become reviewed metric assets in Blender and FreeCAD.
SpatialtemporalAI · Humanoid robotics
MOTIO✳
Agentic Skill Engineering for
Reality-Grounded Humanoid
Loco-Manipulation
A human video. An agent’s context.
A growing world of executable skills.
Humanoid skills · Simulation to realityScroll to discover ↓
From intent to physical interaction
Watch skills take shape.
Ordinary human video becomes task context. Agents build robot-native programs, test them in physics, and refine them through multimodal feedback and human review.
From watching a human
to engineering a skill.
MOTIO brings human-action semantics, metric 3D assets, reusable robot skills, prior cases, and motion references into one agent context.
The agent reasons about stages and contacts, writes executable interactions, and learns from physical feedback—without task-specific teleoperation or robot demonstration collection.
01 / Physical G1 executions
Every step.
Start to finish.
G1 sits, draws, carries a bottle to the bin, and opens a door to transport a cart through the office.
Physical G1 execution24.74 s
Sit. Settle. Rise.
G1 sits on a black office chair and rises to standing.
Physical G1 execution101.75 s
Grasp → carry → dispose.
G1 grasps a bottle, carries it across the office, and places it in the bin.
Physical G1 execution98.41 s
Open the door → transport the cart.
G1 opens the door and guides a cart through the doorway.
Physical G1 execution28.26 s
Physical mouse. Digital drawing.
G1 operates a mouse to draw on screen.
02 / A diverse skill library
Beyond the tabletop.
Locomotion, dynamic balance, collaboration, mechanisms, tools, and deformable materials. Different interactions, one skill-engineering workflow.
Pan tossing
Wrist motion launches and catches the ingredients.
Open the refrigerator
Grasp the handle and pull the door open.
Type a program. Draw a heart.
Physical key presses produce a drawing on screen.
Cucumber slicing
Stabilize the cucumber and guide repeated cuts.
Two humanoids, one sofa
Coordinated contact and large-object transport.
Escalator traversal
Stepping, support, and handrail interaction.
03 / The method
Context. Physics.
Iteration.
Two connected stages turn reality-grounded assets and human action context into executable humanoid skills.
In-context skill engineering
Human video, robot skills, prior cases, and motion priors guide staged programs and contact decisions.
Execute and refine
MuJoCo execution, state measurements, and multi-frame visual feedback guide iteration and human validation.
Agentic 3D · From product geometry to physical interaction ↗
Blender covers broad product modeling; FreeCAD supports dimensioned engineering interfaces. Human asset review connects geometry to the task.
04 / Interactive 3D
Step into a skill.
Explore the robots, objects, and contacts behind MOTIO’s skills. Orbit and zoom to inspect each robot–object configuration.
MOTIO / 3D skill explorer
Loading 3D scene…
Drag to orbit · scroll to zoom · right-drag to pan
05 / Real2Sim2Real
The same room.
A prepared skill.
A scanned office connects appearance and metric geometry to interactive simulation. Layout and object variation prepare the bottle-to-bin task for physical G1 execution.
Real office → interactive world
Appearance. Geometry.
Physical interaction.
3D Gaussian splatting, metric point clouds, and a MuJoCo scene connect the room to skill development. RGB-D, LiDAR localization, and SAM3 support physical deployment.
31/31full-task layout executions
in simulation
in simulation
44/44object-property settings
with a complete grasp window
with a complete grasp window
06 / From engineered skills to a shared policy
Build skills.
Reuse their experience.
Skill executions train one shared policy through BC/DAgger, without task-specific RL.
18tasks
One policy. Many skills.
One shared policy completes 18/18 tasks in development and 18/18 in repeat evaluation.
Reusable skill experience supports coordinated whole-body behavior across tasks.
Median agent-turn time
Across 33 human-validated skills.
Median request to approval
From task request to human validation.
Lower thread residual
Screw-tightening refinement: 269.44 → 64.58 μm.
MOTIO / Paper and resources
Agentic Skill Engineering for
Reality-Grounded Humanoid
Loco-Manipulation
SpatialtemporalAI