Wrong assumptions
Open VLAs expect continuous joints and standard datasets. TonyPi has action groups and a gripper. Nothing fits out of the box.
~/bruhos $ ./wake --body tonypi --lang en
bruhOS
The skill OS that stands next to the robot.
It hears you, runs skills, gets back up, and leaves a trace.
open brains in → skills out → control stays on the body
NVIDIA · LeRobot · open VLMs plug in as planners & data engines
STATE: WALKING
GAIT 1.4Hz · IMU OK · SKILL walk_toward
BODY: TONYPI · RUNTIME: PI-5
failure is a first-class citizen — watch him
fall → stand_up() → keep going
/01
Open brains have nowhere to stand on a real body.
The missing piece isn't another model. GR00T, Cosmos, Isaac Lab, LeRobot, openpi, Qwen-VL — the brains are already open. They plan, simulate, emit action chunks. Meanwhile the TonyPi side is loose action groups, cloud chat, and demos nobody can reproduce.
Two ends that don't connect. Three kinds of waste:
Wrong altitude
Big models can't run on a Pi 5. From the cloud, they hallucinate servos. Either way the body gets hurt.
Wrong ending
A successful run leaves no standard trajectory. No eval, no LeRobot re-ingest, no chain. The data dies at the terminal.
bruhOS plugs the brains that are already open into this body — without rewriting their training code. Runtime owns observation, skills, recovery, trace. The Brain is swappable. The body interface never changes.
/02
One command through the loop.
One command through the loop.
Zero hallucinated servos.
"Put the red one in the box."
EARS
English in
EN-first ASR / TTS / prompts. Short commands, no mixed-language guessing on the hot path.
BRAIN — off the loop
Plans, only plans
Pluggable adapters: OpenAI-compatible API, local Qwen-VL on vLLM, LeRobot policy svc, GR00T System 2. Same schema in, skill calls out.
RUNTIME — Pi 5, on body
Owns the body
observe · dispatch · recover · trace. Compressed observation JSON, whitelisted skills only. stop never waits on a cloud or a block.
BODY + TRACE
Executes & remembers
TonyPi action groups run the skill. Every run writes a LeRobot-aligned, hashable trace — sim and real share one schema.
→ ["detect", "walk_toward(red_block)", "grab", "walk_toward(box)", "release"]
⊘ FORBIDDEN: GR00T action chunk → 16 servos. Adapters are reused — other people's joint definitions are not.
/03
The unit of this system is a skill,
The unit of this system is a skill,
not a chat reply.
- 00look_atorient head / camera to target
- 01detectYOLO / MediaPipe on-body, JSON out
- 02walk_towardclosed-loop approach, no blind steps
- 03follow_linevision-servoed line tracking
- 04grabgripper close, verify hold
- 05releasegripper open, confirm drop
- 06kickballistic action group
- 07wavesocial gesture, EN voice sync
- 08stand_upRECOVERY — a fall resumes the task, it doesn't end it
- 09stopHARD-WIRED — runs on the Runtime alone, always
open models must emit this schema. the runtime is never rewritten to fit a checkpoint's action dims.
/04
What actually changes
every open model gets its own throwaway demo
one skill JSON + one observation schema
the model drives servos → hallucination → the robot eats floor
whitelisted skills only, nothing else moves
raw video shoved into an LLM while the body is still moving
observation compressed on-body, then decided
a fall means the task is dead
recover, then finish the job
mixed-language ASR chaos
EN-first speech, end to end
execution ends as a video clip
LeRobot-compatible trace — simulate it, hash it, chain it
/05
Standing on the open stack.
Standing on the open stack.
Not rewriting it. Not training a base model.
● NOW / v0.1
- Hugging Face LeRobot
- traces & teach sessions export as LeRobot Datasets — swap arms and policies later
- Qwen2.5/3-VL
- instruction + observation JSON → skill calls, local on vLLM
- OpenAI-compatible API
- same schema; swapping the endpoint is the whole migration
- Hiwonder TonyPi
- action groups, camera, IMU, RPC, voice box
- YOLO / MediaPipe / OpenCV
- on-body perception — raw video is never the only input
◐ NVIDIA / wired v0.1 · paying off v0.2+
none of this runs on a Pi. workstation / cluster side — where it belongs.
- Isaac Sim + Isaac Lab
- digital twin of the tabletop task; planning & recovery tested before hardware
- Lab-Arena + LeRobot EnvHub
- the same 20 English prompts regress in sim — brain changes stop breaking robots
- Isaac GR00T N1.7
- System 2 only — task decomposition. Its DiT action head never touches servos
- Cosmos 3
- generated occlusions, shoves, displacements — failure data you can't film enough of
- Isaac Teleop · Jetson Thor
- reserved for the arm era and future compute
○ WHEN EARNED
- openpi π0 / π0.5
- System 1 for the arm, once NexArm / SO-ARM101 lands
- SmolVLA
- light language-conditioned policy on the arm — never full-body on TonyPi
- APXInf
- edge inference engine, Jetson era
- AgiBot World / GO-1
- research reference only — NC licenses stay out of default product weights
/06
Execution == data
v0.1
Local trace
Non-negotiable. Every task writes instruction, skills, observations, outcome.
v0.1→0.2
LeRobot export
Schema aligned now, packaged next — community training tools eat it directly.
v0.2
Cosmos / Isaac loop-back
Synthetic failure cases augment real-hardware acceptance. Never replace it.
v0.3
Hash + ChainSink
On-chain goes trace_root + version hashes of models and skill packs. Not video. Never control.
// HARD RULE
The control loop never blocks on block production or Cosmos inference. stop goes through the Runtime — and only the Runtime.
/07
Roadmap — keep scrolling →
The English skill OS
TonyPi + Qwen-VL / API brains, LeRobot schema, local trace. Ships with a 20-prompt real-hardware eval — completion rate is reported, not vibed.
TONYPI · QWEN-VL · TRACEChange the brain without risking the robot
Isaac Lab regression on the same 20 prompts. Cosmos fills the failure data. GR00T S2 adapter goes live as an option.
ISAAC LAB · COSMOS · GR00T S2Walk, see, speak / grasp — split
Two machines: TonyPi runs skills, the arm runs π0.5 / SmolVLA. ChainSink can be switched on.
DUAL BODY · π0.5 · CHAINSINKEdge VLA, when the hardware earns it
Jetson-class compute first, then on-device VLA talk. Still no GR00T full-body action head on a Pi 5. Ever.
JETSON · APXINF
/08
Success is written down before the demo.
- C1Swapping the Brain adapter (API ↔ Qwen-VL ↔ GR00T S2 stub) changes zero skill names.
- C2Completion rate on 20 English real-hardware prompts is reportable — the sim suite runs the same prompts.
- C3Shove-recovery rate is accounted separately. Getting back up is a metric, not a blooper reel.
- C4Every task exports a hash; fields map onto LeRobot; killing Isaac or the chain module never affects stop.
- ✕Chinese-default ASR / an unregistered skill / GR00T joints straight to servos — any of these = test failure. Full stop.
runs like
your bruhhh.
bruhOS exists because open brains have nowhere to stand on a real body. It plugs NVIDIA, LeRobot, and open VLMs into TonyPi as planners and data engines — while the robot itself only runs skills, speaks English, gets back up, and leaves a trace.
[ read the product doc ]
bruhOS.ai