remote-training

github:Positronic-Robotics/positronic · first committed 2026-01-05 · source · badge

Manages remote training infrastructure on Nebius VMs. Use for building/pushing Docker images, starting/stopping VM machines (train, train2, train3), running training jobs, dataset generation, and starting inference servers.

Skill Health: not yet scored

Listed, not yet scored. This skill has no public model test yet. It is in the scoring queue; submitted skills go first, then the most-installed.

Version history

33 commits to this file on GitHub (newest first).

2026-10-27 · commit cf5efa2 · Build `positro/openpi-base` in CI from a pinned openpi fork commit (#1049)
2026-10-25 · commit 7f6bb88 · Report and solve every Franka arm at the `gripper` site (#1021)
2026-10-25 · commit 1d2b21e · State the frame of each pose signal as `ee_frames` (#1020)
2026-09-28 · commit 4005412 · Add a keepalive call, and warm each server when it loads (#754)
2026-09-25 · commit 4fec81f · Serve one checkpoint per server and remove checkpoint switching (#816)
2026-09-22 · commit 1c2c410 · Reach an offboard server over a Unix socket, on the `websocket_unix` wire (#735)
2026-09-21 · commit be7f03c · Ship the client side of both wires as the `positronic-wire` distribution (#779)
2026-08-28 · commit 64de087 · Let the episode call carry the session that runs it (#690)
2026-08-11 · commit 95b929c · Serve inference endpoints over Nebius managed HTTPS URLs with token auth (#565)
2026-08-09 · commit 65af747 · Declare the EE frame every openpi deployment serves in (#583)
2026-07-26 · commit 29285cf · Define a policy as one pipe terminated by a `ModelSource` (#497)
2026-07-21 · commit d1f7aef · docs: genericize hardcoded `vertix` home path in remote-docker docs (#492)
2026-06-10 · commit cfc2aeb · Add `positronic eval run` with first-class `Task` + `Eval` (#420)
2026-05-28 · commit ec4c28b · Pin `uv.lock` + add `recording_dir` to remote policies (#408)
2026-05-25 · commit 498ef5e · Reorganize PhAIL configs by release version (#406)
2026-05-19 · commit ebe6cef · Teach `remote-training` skill local desktop/notebook serving (#399)
2026-05-18 · commit 2888743 · Serverless endpoint auto-shutdown + remote-training skill for Nebius Serverless (#392)
2026-03-18 · commit bef7427 · Fix SmolVLA training: LR decay, deferred run metadata (#350)
2026-03-18 · commit ecc0b2d · Add `phail` preset to lerobot 0.4.x server (#353)
2026-03-14 · commit a34fb05 · Fix SmolVLA training: LR decay, batch size, full finetune config (#346)
2026-03-05 · commit a73da96 · Update docs and skill for lerobot 0.3.3/0.4.x service split (#336)
2026-03-04 · commit 03cd423 · Add production server configs, weighted remote policy, and eval UI object selection (#334)
2026-03-04 · commit 7351405 · Add lerobot 0.4.x vendor (SmolVLA, ACT, Diffusion) (#328)
2026-03-02 · commit cf925e4 · Update remote-training skill with S3 convention and current state (#331)
2026-02-27 · commit 6a810bf · Make `lerobot` an optional dependency (#324)
2026-02-24 · commit 77857f6 · Update docs: s3 output and conversion commands (#319)
2026-02-24 · commit 88f94ef · Remove `--fps` from dataset conversion examples (#317)
2026-02-23 · commit 3ddcf8d · Composable Codec with `|` and `&` operators (#312)
2026-02-20 · commit e06535a · Composable Codec, action timestamps, GR00T merge (#307)
2026-02-17 · commit ab41e86 · Add sim eval end-to-end section to remote-training skill (#306)
2026-01-14 · commit 705402f · Refactor dataset configs into `cfg/ds/` with public/internal separation (#277)
2026-01-13 · commit 5ce0dc5 · Refactor GR00T server to match lerobot server API (#273)
2026-01-05 · commit 4261b1a · 6D rotation representation with unified GR00T action/observation encoding (#268)

Archived snapshots