LeRobot · Proof of Concept Engagements

We run the PoC.
On your task.

A Quintinity LeRobot engagement takes a manufacturing task from concept to trained, deployed policy in four to eight weeks. We bring the hardware, the stack, and the method. You bring the task and the team.

Scope· Setup· Collect· Train· Evaluate· Deploy

Six phases. One trained policy.

Every engagement follows the same method. The phases are fixed; the task is yours. We move at the speed of the data — the only bottleneck is how fast we can collect good demonstrations.

01 · Scope
Task definition & feasibility
We start at your facility. We define the task precisely: what the arm must pick up, where it must place it, what counts as success. We assess workspace geometry, lighting, object variation, and edge cases. We confirm that imitation learning is the right method for this task — or we tell you it isn't.
Site visit (0.5 days)
Task specification document
Hardware selection (SO-101 vs B601)
02 · Setup
Hardware, calibration & environment
We bring the arm, mount it at the workstation, configure cameras, and run the calibration routine. Leader and follower arms are zeroed against each other. The data collection pipeline is tested end to end before a single demonstration is recorded.
lerobot-calibrate (joint zeroing)
Dual-camera mount (wrist + overhead)
NVIDIA Jetson Orin or DGX Spark
End-to-end recording test
03 · Collect
Teleoperation & dataset
A Quintinity engineer or trained operator runs the leader arm through the task repeatedly. We target 100–200 demonstrations, discarding failed attempts. The dataset — joint states, gripper, camera frames at 30fps — is pushed to the Hugging Face Hub.
lerobot.scripts.control_robot
Hugging Face Hub (dataset repo)
100–200 valid episodes
Episode QA (reject rate logged)
04 · Train
Policy training
We train an ACT (Action Chunking with Transformers) or diffusion policy on the collected dataset. Training runs on our DGX Spark or a cloud GPU node. A 52M-parameter ACT model converges in 4–8 hours. Checkpoints are pushed to the Hugging Face Hub.
lerobot.scripts.train
ACT policy (52M params, default)
Diffusion policy (if task demands)
DGX Spark / RTX 3090 / A100 cloud
Weights & Biases (training metrics)
05 · Evaluate
Rollout & iteration
We load the policy checkpoint to the follower arm and run rollouts against the real task. We measure success rate, failure mode taxonomy, and generalisation to object variation. Each failure mode drives a targeted collection run to fill the gap.
lerobot.scripts.eval
Failure mode log (Notion / markdown)
Targeted top-up data collection
Retrain cycle (repeat until threshold)
06 · Deploy
Handoff & documentation
The final policy runs autonomously on the follower arm. We document the task specification, dataset, training config, and evaluation results. The dataset and model are published to the Hugging Face Hub. Your team is trained to collect new demonstrations and retrain.
Policy checkpoint (HF Hub, public or private)
Dataset published (CC BY 4.0)
Runbook: re-train & deploy
1-day team training session

What you receive at the end.

Primary
Trained policy checkpoint
A policy that runs autonomously on the follower arm at the agreed success rate. Published to Hugging Face Hub. Reproducible from the dataset.
Primary
Demonstration dataset
100–200+ curated episodes on the Hugging Face Hub. Re-trainable. Extendable. Shareable with the open-source community if you choose.
Primary
Runbook & trained team
Full documentation: task spec, training config, evaluation results, and how to collect more data and retrain without Quintinity's involvement.
Secondary
Failure mode taxonomy
A structured log of every failure mode encountered during evaluation, with root cause and suggested collection strategy for each.
Secondary
Hardware recommendation
Based on the PoC result: whether the task is ready for the reBot B601 production cell, or needs additional development work.
Optional
Published open-source release
With your agreement, we publish the dataset and checkpoint publicly on Hugging Face and contribute to the LeRobot community. ANZ industry on the map.

Hardware we bring.

SO-ARM101 leader + follower pairTeleoperation
reBot Arm B601-DMPre-production cell
NVIDIA Jetson Orin NX 16GBEdge inference
NVIDIA DGX SparkPolicy training
Dual USB cameras (wrist + overhead)Vision input
Workstation mounting hardwareCell setup

Software stack.

Hugging Face LeRobotCore framework
ACT policy (default)Imitation learning
Diffusion PolicyAlt. policy type
Hugging Face HubDataset + model hosting
PyTorch + CUDATraining backend
Weights & BiasesTraining observability
NVIDIA Isaac SimSimulation (optional)

Typical engagement timeline.

Scope
0.5d
Site visit, task definition, go/no-go
Setup
1–2d
Hardware install, calibration, pipeline test
Collect
2–5d
100–200 demonstrations, dataset QA
Train
4–12h
ACT or diffusion policy, checkpoint to HF Hub
Evaluate + iterate
1–2w
Rollouts, failure analysis, top-up collection
Deploy
1d
Handoff, documentation, team training

What we need from you.

A defined, repeatable task. Pick-and-place, sort, insert, inspect — anything a human can demonstrate with both hands.
A stable workspace we can mount the arm at for the duration. Consistent lighting preferred.
One person from your team who knows the task well enough to evaluate whether the policy is succeeding.
Willingness to say "that's not good enough" when the policy fails an edge case. Honest evaluation is the fastest path to a working policy.

What we never promise.

We never promise a specific success rate before we have seen the task and collected data. Anyone who does is guessing.
We never blur "experimental" with "production-grade." If the PoC achieves 80% on a lab task, that is not a production-ready cell.
We never name a client publicly without signed permission.
We never let the hardware manufacturer be the hero. LeRobot, Seeed, NVIDIA — they are ingredients. The engagement is the product.

Bring us your line. We will model it in a week.

One ask per conversation. No deck, no funnel. Tell us the task, the factory, the team. We will tell you whether a LeRobot PoC makes sense — and what it would take.

Book a 15-min call Read the rebuild thesis