E1 · Train for 200 steps
15 minRun the lerobot-train from the panel (with cpu or cuda). Time it.
Done when: ls outputs/train/act_pusht_teste/checkpoints/ shows 000200 and last.
Lesson 5 · 18 min
Hands-on with LeRobot 0.6
LeRobot, from Hugging Face, puts data, policies, and commands for recording, training, and running in one place. In this lesson you train a real ACT policy for a few steps and end up with a checkpoint on disk. No robot and no account needed.
Start the lesson
In this lesson you will
- 1The record → train → eval → rollout cycle
- 2Train an ACT in 200 steps and find the checkpoint
- 3Read the loss in the log without fooling yourself
- 4How long it takes on GPU and on CPU
Goal: install LeRobot 0.6, train a small policy, and understand the record → train → run cycle.
01The LeRobot cycle
Four commands. The middle two don't need a robot. Click each one.
ACT = Action Chunking with Transformers: a small policy that returns chunks of actions, the idea from lesson 2.
Recording and running need the arm. Training and evaluating run on any computer.
02What you need to know before training
Click each step.
1. Installation 0.6
Since 0.6, the base package is lean: to train, install lerobot[training]. For simulation, use extras like [pusht] or [aloha].
Python 3.12 and PyTorch 2.7 or newer. Some options were renamed: eval_freq became env_eval_freq, and --dataset.vcodec became --dataset.rgb_encoder.vcodec.
2. Hardware
The 200 steps in this exercise took 26 s on an NVIDIA GPU (GB10) and 321 s, about 5 minutes, on just the 20-core CPU of the same machine (10/19/2026). A regular laptop will be slower.
Real training is a different scale: the LeRobot docs say training on your own robot takes several hours. The 0.6 hardware guide measures memory and time per policy family.
3. No GPU?
For this exercise, a CPU is enough. For long training runs, 0.6 supports the same lerobot-train in the cloud with --job.target, through HF Jobs, which is billed by usage and requires an account.
You don't need it for the course: everything here runs locally.
4. With the SO-101
The SO-101 is LeRobot's entry-level arm. With the leader and follower pair, you record your own dataset with lerobot-record and train the same way.
LeLab runs the whole flow in the browser, with no command line, and today it only supports the SO-ARM101.
03Train an ACT for 200 steps
Same venv as lesson 4. Swap cuda for cpu if you don't have an NVIDIA GPU.
Terminal · short training run
lerobot-train \ --policy.type=act \ --dataset.repo_id=lerobot/pusht \ --steps=200 \ --save_freq=200 \ --log_freq=50 \ --batch_size=8 \ --policy.device=cuda \ --policy.push_to_hub=false \ --wandb.enable=false \ --output_dir=outputs/train/act_pusht_teste \ --job_name=act_pusht_teste
Terminal · check the checkpoint
ls outputs/train/act_pusht_teste/checkpoints/ ls outputs/train/act_pusht_teste/checkpoints/last/pretrained_model/
With the SO-101 (from the docs, needs the arm)
lerobot-record \ --robot.type=so101_follower \ --robot.port=<follower port> \ --teleop.type=so101_leader \ --teleop.port=<leader port> \ --dataset.repo_id=<your username>/meu-dataset \ --dataset.num_episodes=5 \ --dataset.single_task="Pick up the black cube" \ --dataset.streaming_encoding=true \ --dataset.encoder_threads=2
Example from the repository
INFO ... cfg.steps=200 (200) INFO ... dataset.num_frames=25650 (26K) INFO ... dataset.num_episodes=206 INFO ... num_learnable_params=51588994 (52M) INFO ... step:50 ... loss:14.587 ... INFO ... step:100 ... loss:4.694 ... INFO ... step:150 ... loss:3.878 ... INFO ... step:200 ... loss:3.473 ... mem_gb:0.95 INFO ... Checkpoint policy after step 200 INFO ... End of training $ ls outputs/train/act_pusht_teste/checkpoints/ 000200 last $ ls outputs/train/act_pusht_teste/checkpoints/last/pretrained_model/ config.json model.safetensors policy_postprocessor.json ... train_config.json
LeRobot 0.6.1, NVIDIA GB10 GPU: 26 s total. CPU only: 321 s, final loss 3.536. Lines shortened with "...". The first time, it also downloads the ResNet-18 weights that ACT uses for vision.
- Forgetting
--policy.push_to_hub=false: without it, training stops right at the start withValueError: 'repo_id' argument missing. Please specify it to push the model to the hub.(tested on 10/19/2026): by default it wants to push the policy to the Hub. - Thinking low loss = good robot: the loss measures how well it imitates the dataset. The evaluation in lesson 6 tells you whether the policy works.
--policy.device=cudawithout an NVIDIA GPU: swap it forcpu.
04Exercise: your first checkpoint
About 5 minutes on CPU, under 1 minute on GPU.
E2 · Read the loss
5 minIn the log, find the step:50 and step:200 lines and write down both losses.
Done when: the loss at step 200 is lower than at step 50, and you noted how long the training took on your machine.
E3 · More steps (optional, with a GPU)
15 minRepeat with --steps=5000 --save_freq=5000 --batch_size=32 and a different --output_dir. Keep the checkpoint: lesson 6 compares the two.
Done when: the checkpoints/005000 folder exists (here it took 474 s on the GPU, with a final loss of 0.329).
05Test what you learned
The explanation appears after your answer.
Question 1 of 4
Which commands in the cycle run without a robot?skill
Correct
Training uses a dataset (it can come from the Hub) and evaluating uses simulation. Recording and rolling out need the arm.Not this one
Training uses a dataset (it can come from the Hub) and evaluating uses simulation. Recording and rolling out need the arm.Review: The LeRobot cycleQuestion 2 of 4
Where is the checkpoint after training?
Correct
With --output_dir=outputs/train/act_pusht_teste, the checkpoints go in checkpoints/000200 and checkpoints/last. With push_to_hub=false, nothing is uploaded.Not this one
With --output_dir=outputs/train/act_pusht_teste, the checkpoints go in checkpoints/000200 and checkpoints/last. With push_to_hub=false, nothing is uploaded.Review: Train an ACT for 200 stepsQuestion 3 of 4
The loss dropped from 14.6 to 3.5. What does that guarantee?
Correct
Loss measures the imitation error on the training data. In lesson 6 you see that 200 steps still give 0% success on PushT.Not this one
Loss measures the imitation error on the training data. In lesson 6 you see that 200 steps still give 0% success on PushT.Review: Train an ACT for 200 stepsQuestion 4 of 4
What is the dagger strategy in lerobot-rollout for?
Correct
Run, correct, and fine-tune again: the corrections are tagged with intervention and go into the next training run. This is the link to lesson 7.Not this one
Run, correct, and fine-tune again: the corrections are tagged with intervention and go into the next training run. This is the link to lesson 7.Review: The LeRobot cycleQuiz complete
You got {a} of {n} right. Get them all right and the lesson is marked as complete.Pick an option. The explanation appears after your attempt.
06Summary
Four commandsrecord, train, eval, rollout. Train and eval don't need a robot.
Installation 0.6lerobot[training], Python 3.12, push_to_hub=false to train without an account.
Real time200 steps: 26 s on GPU, 5 min on CPU here. Real training: hours.