Knowledge insulation (KI) is a training recipe that keeps action learning from damaging a vision-language-action model's pretrained backbone. Discrete FAST action tokens train the backbone, while a stop-gradient keeps a flow-matching action expert's gradients out of itSource 1.

A strip of frames from a robot camera showing a mobile robot clearing dishes, cups and trash from a table into a bin.
Table bussing is the task behind KI's training-cost result: a many-embodiment π0 generalist needed 7.5 times the training steps of π0.5 with KI to reach similar performance. Frames of the task from the FAST paper.
Paper figure: Karl Pertsch et al. / arXiv. Licence: CC BY 4.0.

Turn off "Protect language skills" to remove the stop-gradient.

Knowledge insulation
Shield the language brain while teaching the robot to move
Adjust more

Backbone supervision: FAST-token cross-entropy only

Shut valves stop movement lessons, so they can't overwrite what the model knowsbackward · 8 of 8 trays · valves shut
Wide column: understands words and picturesSlim column: moves the arm
Follows instructions (illustrative, not measured)92 / 100
Teaching movement can wear away an AI's grasp of language; walling it off keeps that grasp, and training needed 7.5 times fewer steps than π0.
How this was made

The model is a pi0-style vision-language-action model laid out as a mixture of transformers: a 2B language-model backbone and a 300M flow-matching action expert. The drawing makes each a column of 8 trays, a schematic layer count, with cross-sections in the ratio of their sizes, and joins them at every tray through a one-way check valve. On the forward pass the expert attends sideways into the backbone's activations at every layer, so the valves open. On the backward pass the expert's flow-matching loss sends gradients back down its own stack. Knowledge Insulation puts a stop-gradient at the interface (every valve shut), so those gradients never reach the backbone; the backbone instead learns actions from the cross-entropy loss on discrete FAST action tokens, the feed into its head. Turn the protection off and the expert's gradients cross into the backbone at every layer they reach.

The 0 to 100 “follows instructions” score is illustrative: it stands in for the paper's spoon-and-trash example, where a model trained without insulation stops following the instruction, and is not a published curve. The measured result is separate: 7.5x fewer training steps for the π0.5 + KI generalist than π0, at similar table-bussing performance. Source: Driess et al. 2025

Backward pass at depth 8 of 8 keeps backbone supervision on FAST-token cross-entropy only, language following at 92 of 100, and the separately reported 7.5x fewer training steps for the π0.5 + KI generalist versus π0 at similar table-bussing performance; the stop-gradient is on so expert gradients stay inside the action expert.

Current MoT pass
pass
backward
depth
8 / 8
stop-gradient
on
supervision
FAST-token cross-entropy only
language score
92 / 100
Schematic with 8 layers per stack; the instruction score is illustrative. Training-step result from the Knowledge Insulation paper.
7.5x fewer
Training steps
π0.5 + KI generalist on table bussing, versus π0
3B params
VLM backbone
complete VLM in the research-note diagram
300M params
Action expert
paper configuration, trained with flow matching
Continuous
Runtime output
FAST action tokens provide training supervision

Origins

RT-2 tokenizes each continuous action dimension into 256 uniform bins, as RT-1 didSource 7, and Physical Intelligence's 2025 KI note judges such fixed-bin tokens unsuitable for high-frequency, precise or fluent motionsSource 2. π0 adds a continuous action expert to a pretrained VLMSource 4. π0.5 trained 280k steps on discrete tokens, then post-trained a randomly initialized flow-matching expertSource 8; KI makes this a single stageSource 1, called π0.5 + KISource 2.

Naive joint training

Fine-tuning a pretrained network through a random head distorts its features and raises out-of-distribution errorSource 10. The KI paper finds the same in VLAs: gradients from a new action expert degrade the backbone, slowing training and weakening language following, so that π0, told to bus a spoon into the bin, grasps trashSource 1.

ChatVLA calls this spurious forgetting, robot training overwriting visual-text alignmentsSource 17. Between robot tasks, though, pretrained VLAs resist forgetting, and simple experience replay sometimes removes itSource 27.

Recipe

  1. The action expert trains with flow matchingSource 12 and generates continuous action chunks at inferenceSource 1.
  2. The backbone learns from discrete FAST action tokens plus vision-language and robot-planning data. FAST applies a DCT to each action dimension over time, rounds the coefficients (the lossy step) and compresses the sequence with byte-pair encodingSource 3.
  3. Stop-gradient on the backbone keys and values the expert reads keeps the forward path and cuts the backward one. The expert attends to the image, language and state prefix and to itself, never to FAST tokens, and the backbone never attends to itSource 1.

Training both outputs costs about 20% more compute. The paper's backbone is a 2B language model with a 300M expertSource 1; the note labels the complete VLM 3BSource 2. FAST tokens are π0.5 + KI's training representation and continuous flow actions its runtime outputSource 1, so a comparison matrix needs both columns.

Limits and alternatives

Stopping expert gradients does not adapt the backbone: with the backbone frozen, the paper's shirt-folding policy scores 0%. FAST supervision supplies that signal, and vision-language co-training can give plain joint training good language following without stop-gradientSource 1.

GR00T N1 freezes the language part of its VLMSource 13. NVIDIA reports that N1.5, which froze the whole VLM among other changes, raised real GR-1 language following from 46.6% to 93.3%Source 14, and N1.6 unfreezes the top 4 layersSource 15.

BLIP-2 bridges frozen image encoders and language modelsSource 11; VLM2VLA writes actions as language and trains only LoRA adaptersSource 19, and UAM keeps over 95% of its VLM's multimodal capability with no freezing, gradient stopping or co-trainingSource 25. Figure's Helix instead backpropagates from its fast policy into the VLM through their shared latent vectorSource 16, and HybridVLA trains discrete and diffusion action prediction togetherSource 18.

Results

A many-embodiment π0 generalist needs 7.5 times the training steps of π0.5 + KI to reach similar table-bussing performance, and single-embodiment π0-FAST specialists need twice the wall-clock time from slow inference. Table bussing is scored at ten episodes per task and policy. When a mobile robot moves unseen objects into an open drawer, web-data co-training boosts generalization most, and language following improves but stays imperfectSource 1 Source 2.

In third-party fine-tuning, π0.5 with KI stays near zero on out-of-distribution task perturbations, so KI alone does not fix out-of-distribution language generalizationSource 23. A reimplemented co-training plus KI baseline trailed standard behavior cloning on mean LIBERO-PRO and LIBERO-Plus successSource 24, though the dual FAST and flow objective mitigates forgetting more than flow aloneSource 26. Across seven VLAs, VQA co-training goes with better knowledge retentionSource 28, and in Toyota Research Institute's 89-policy study, training only on robot data cost the VLM backbone its ability to generate languageSource 29.

Later models

openpi's public π0.5 checkpoints were trained with KI, but the repository serves only the flow-matching headSource 9, so outside groups reimplemented the FAST loss and stop-gradientSource 26. LAP-3B blocks expert gradients and runs at 25 Hz on an RTX 4090Source 20, and LabVLA applies KI to a FAST-pretrained Qwen3-VL-4BSource 21. Dexmal and StepFun's DM0 (February 2026) adopts a hybrid gradient strategy "inspired by Knowledge Insulation"Source 30, and a 2026 review sorts implementations into gradient stopping, as in π0.5-KI and π0.6, and inpaintingSource 22.

An overhead view of two robot arms folding a flat cardboard sheet into a box on a table.
π0.6, whose card lists a Gemma3 4B backbone with an insulated expert of about 860M parameters, building a box. Physical Intelligence used this task to test training-time real-time chunking.
Photo: Kevin Black et al. / arXiv. Licence: CC BY 4.0.

The π0.6 card lists a Gemma3 4B backbone predicting FAST tokens and an approximately 860M expert whose gradients stay out of the backbone; a chunk takes 63 ms on one H100 with five denoising steps and three camerasSource 5. π0.7 keeps a Gemma3-based VLM and a flow-matching expert under FAST supervision with expert gradients stoppedSource 6.