Training
SFT or RL, on your compute.
Fine-tune on a dataset (SFT), or run RL against a query set with graders as the reward. Common open stacks, jobs dispatched to the customer's cluster, results verified on real hardware.
Job
SFTRL
artha train.create(
mode = "rl", # or "sft"
base = "pi0.5", # openpi checkpoint
query_set = "kitchen_manip_v1", # tasks + rubrics (Evaluation)
reward = "graders/load_dishwasher@v4",
rollouts = "sim:kitchen_counter_v3 × Humanoid-G1",
cotrain = "ds_kitchen_real_v1", # real demos mixed in
method = "advantage-conditioned", # RECAP-style
compute = "customer://k8s/gpu-pool-a",
)
# SFT variant
artha train.create(mode="sft", base="gr00t-n1.7",
dataset="ds_kitchen_h2r_v2", compute="customer://slurm/a100")
Supported stacks
openpiπ0 / π0.5
Isaac-GR00TGR00T N1.x
LeRobotACT, Diffusion, SmolVLA
Isaac Lab / MJXsim-side RL
Your compute
Kubernetesconnected
Slurmconnected
Artha Labs-hostedoptional
Weights and data stay in the customer's environment. Artha Labs ships the containers, configs and grader services.
RL loop: train, then verify on real hardware
Sim success per RL epoch · π0.5 fine-tune (RECAP-style) · batch #5512
epoch 1epoch 6 (proj.)
Real-world check
65% ± 5
real success, 95% CI, n = 350
sim success71%
sim-to-real gap−6 ± 5 pts
Methodology on Evaluation →
Export: what the customer receives
artha dataset.export #5512 \
--format lerobot \
--include sim,real_eval,grader_traces \
--filter physics_consistency=1.0
| Artifact | Count |
|---|
| RL-ready trajectories | 160,512 |
| Reward-hacks excluded (12%) | 21,888 |
| Real-eval episodes | 350 |
| Grader traces (auditable) | all |
The deliverable: a LeRobot-format dataset + auditable grader traces + a real-eval validation report, or the fine-tuned checkpoint itself. Neutral to whose model consumes it.
Mock only — illustrative data.