VLA Gallery / All models / π family

학습시 RTC

Training-time RTC (Training-time RTC, π0.6)

Physical Intelligence · 2025-12

기타π familyBased on π0.6hybrid

Simulates inference latency during training (uniformly sampled 0–10 steps on hardware) to learn action-prefix conditioning, so there is no extra compute from inference-time RTC.

Paper

Training-Time Action Conditioning for Efficient Real-Time ChunkingarXiv 2512.05964

Overview

Category

π family

Type

기타

Released

2025-12

Organization

Physical Intelligence

Open weights

No

Parameters

Based on π0.6 (no separate figures)

Architecture

Backbone LLM

Gemma 3 4B (π0.6)

Vision encoder

SigLIP 400M (π0.6)

Action head

π0.6 flow expert (trained with latency simulation) (hybrid)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

π0.6 as is (no architecture·runtime change); simulates inference latency during training and conditions on the already-executed action prefix

Training

Pretraining accelerators

N/A (based on π0.6)

Pretraining duration

N/A

Pretraining steps & batch

N/A

GPU/TPU-hours

Not disclosed

Fine-tuning

8,000 steps·batch 512 per task (hardware·time not disclosed)

Data type

Real-robot teleoperation (per-task fine-tuning data)

Data scale

Not disclosed

Main datasets

Per-task data (not disclosed)

Tasks & evaluation

Box assembly, espresso making (π0.6); simulation comparison under matched training compute

Robot

π0.6 real-robot platform (50 Hz robot)

Inference

Inference device & speed

Remote H100, 5 denoising steps: training-time RTC 108 ms (d≈5) vs inference-time RTC 135 ms (d≈7)

Inference memory

Not disclosed

Weights bf16 (calc.)

Not disclosed

Where it runs

Remote server

Reinforcement learning & base

Reinforcement learning

Not used · Supervised fine-tuning with simulated inference latency

Base control

Fixed · No mention of base control (box assembly·espresso tasks)

Notes

Key features

Simulates inference latency during training (uniformly sampled 0–10 steps on hardware) to learn action-prefix conditioning, so there is no extra compute from inference-time RTC. On box assembly and espresso it matched inference-time RTC with lower latency.

Notes

Type is Other since it is a fine-tuning technique based on π0.6. Supports up to 200 ms latency on a 50 Hz robot. Simulation samples latency {0..4} with exponentially decaying weights.

Verification

Confirmed