One-Shot Imitation Learning Software
A deterministic
one-shot engine.
For hardware manufacturers and software integrators, the core bottleneck to scaling is the time, cost, and fragility of robot deployment. Lili-o replaces massive, data-heavy AI pipelines with a deterministic, one-shot execution engine.
Record a single demonstration. We capture the trajectory and the scene geometry, then re-derive execution for whatever the environment throws at it — adapting to pose variation and semantically similar objects without ever retraining.
Two usage for teams.
Accelerate Sales.
Stop wasting weeks programming custom proof-of-concept demos. Record one demonstration on-site via kinesthetic teaching or VR, and Lili-o instantly converts it into a repeatable live demo that adapts to real-world clutter, pose shifts, and new object geometry — proving your hardware's value to prospects on the spot.
Fail-Safe Redundancy.
Protect your AI brain from edge-case failures. When end-to-end foundation models or RL policies collapse on out-of-distribution shifts, Lili-o acts as a zero-shot, deterministic fallback — mapping trajectories to any shared Cartesian task space and running entirely on local embedded hardware with zero cloud dependencies.
It's a data treadmill.
Imitation learning and VLAs deliver impressive results in controlled settings — but they require hundreds to thousands of demonstrations per task, per robot, per environment, and collapse the moment objects or scenes drift out of distribution. RL offers a complementary path, but sim-to-real quality is tightly coupled to physics identification, and learned policies stay brittle to distribution shift. Every environment change resets the treadmill.
demonstrations per task configuration to bring a Diffusion Policy to production quality.
of demonstration data required to pretrain a frontier-scale VLA (GEN-0).
Learned once. Executed repeatedly.
For repetitive operations in structurally stable environments, you don't need a policy that generalises across everything. You need one that executes reliably, adapts to pose variation, and transfers to semantically similar objects without retraining.
Demonstrate once
Via VR/AR teleoperation or kinesthetic teaching, we capture two things at once: the full 6-DOF gripper trajectory in the robot base frame, and a coloured 3D point cloud of the scene. Each atomic interaction — grasp, place — is its own demonstration unit.
Register at execution
We segment the current scene and align the live point cloud against the reference using geometric registration, then project the updated object pose into the robot base frame.
Recompute & run
The full trajectory is recomputed for the new pose and executed. No retraining. No fine-tuning. No additional data collection — a task learned once, executed repeatedly across varying poses, positions, and lighting.
What it means operationally.
- Semantic transfer to new object instances without retraining — covering objects that share functional geometry with demonstrated ones.
- Runs on an embedded GPU for perception and a standard CPU for trajectory execution. No GPU server, no cloud.
- Deploys across robot embodiments without embodiment-specific retraining, given a shared Cartesian task space.
The system is deliberately scoped: it performs at its best on repetitive tasks where the structural layout is stable but object pose, robot configuration, and object instances vary — a large fraction of commercially relevant manipulation deployments today.