Skip to content
RFL_GLOBAL
中文
  • WAVE 5 // POST-TRAINING
  • PYGMALION
  • FROZEN VLA UPGRADE

CHRYSALIS

THE METAMORPHOSIS

Improve existing VLA models WITHOUT retraining. CHRYSALIS transfers dynamics priors from skill-compositional world models into frozen VLAs through two-stage alignment and residual correction. VLA weights stay frozen — CHRYSALIS adds a lightweight correction layer that transforms base model outputs into expert-level trajectories.

MODULE STATUS: RESEARCH

No Retraining Needed

0-TRAIN

DIVISION
ANIMA
WAVE
W5
DOMAIN
MANIPULATION
WAVE 5 // ANIMA SUITE
MANIPULATION — LATENT POST-TRAINING FOR VLA
CHRYSALIS // W5 // 012/079
01THE CHALLENGE

RETRAINING VLAS IS PROHIBITIVELY EXPENSIVE

Vision-Language-Action models are massive. Training them from scratch costs millions in compute. Fine-tuning risks catastrophic forgetting — the model improves on new tasks but loses previously learned capabilities. Every deployment environment is different, but retraining for each is impossible at scale.

What's needed is a way to improve VLAs after deployment without touching their weights. A post-training correction layer that absorbs dynamics priors from world models and skill libraries — turning a generic VLA into a domain-specific expert without the cost of retraining. No one has done this cleanly.

02THE SOLUTION

WHAT CHRYSALIS DELIVERS

CHRYSALIS adds a lightweight residual correction layer on top of frozen VLA outputs. It transfers dynamics priors from skill-compositional world models through a two-stage alignment process — first aligning latent representations, then learning residual corrections to VLA action outputs.

CAPABILITIES

  • Frozen VLA improvement: enhance existing VLA models without modifying their weights
  • Two-stage alignment: latent representation alignment followed by action residual correction
  • Skill decomposition: break complex tasks into composable skill primitives from world models
  • World model coupling: transfer dynamics priors from trained world models into correction layers
  • Catastrophic forgetting prevention: base VLA capabilities preserved — corrections are additive
  • Lightweight deployment: correction layer adds minimal compute overhead to base VLA inference
03ENGINEERING

WHY THIS IS HARD

Latent post-training for frozen VLAs is an unsolved problem requiring:

  1. 01Representation alignment: world model latent space and VLA latent space must be bridged without shared training
  2. 02Residual learning: corrections must be small enough to preserve base model behavior but large enough to meaningfully improve performance
  3. 03Skill composition: decomposing world model knowledge into reusable skill primitives that transfer across tasks
  4. 04Stability guarantees: correction layer must never degrade base VLA performance — only improve or be neutral
  5. 05Efficient inference: added correction layer must not significantly increase latency or memory requirements

CHRYSALIS achieves this through careful two-stage alignment: first bridge the latent spaces, then learn residual corrections. The butterfly emerges — better than before, with all original capabilities intact.

04BENCHMARKS

KEY PERFORMANCE METRICS

Target metrics for the CHRYSALIS architecture:

KEY PERFORMANCE METRICS
METRICVALUEDETAIL
No Retraining NeededCONFIRMEDVLA weights stay completely frozen — correction layer is the only trainable component
Skill DecompositionACTIVEComplex tasks broken into composable primitives from world model dynamics
World Model CouplingENABLEDDynamics priors transferred from skill-compositional world models
Frozen VLA ImprovementTARGETMeasurable performance uplift on manipulation benchmarks without weight modification
05BUILD STATUS

WHAT'S BUILT TODAY

1/5 COMPONENTS COMPLETE
WHAT'S BUILT TODAY
COMPONENTSTATUSNOTES
PRD & Architecture DesignCOMPLETEFull product requirements document and system architecture
Core modelsIN PROGRESSWorld model + correction layer architecture — design phase
Two-Stage AlignmentPLANNEDLatent space bridging + residual correction learning
Residual Correction LayerPLANNEDLightweight additive corrections to frozen VLA outputs
API layerPLANNEDREST + gRPC endpoints — pending core model implementation
06APPLICATIONS

WHERE CHRYSALIS DEPLOYS

  • APP_01

    POST-TRAINING UPGRADES

    Improve deployed VLA models in the field without retraining — ship correction layers as lightweight patches.

  • APP_02

    CONTINUOUS IMPROVEMENT

    Iteratively refine VLA performance by training new correction layers on domain-specific world model data.

  • APP_03

    CATASTROPHIC FORGETTING PREVENTION

    Safely add new capabilities without losing existing skills — base VLA remains untouched, corrections are additive.

07TECHNOLOGY

UNDER THE HOOD

FOUNDATION: LATENT POST-TRAINING

  • Frozen VLA backbone: base model weights never modified during post-training
  • Skill-compositional world models: dynamics priors decomposed into reusable primitives
  • Two-stage alignment: (1) bridge latent spaces, (2) learn residual action corrections
  • Additive correction: outputs = VLA_output + correction_layer(VLA_latents, world_model_priors)

KEY INNOVATION

CHRYSALIS treats VLA improvement as a post-training problem. Instead of retraining the butterfly, add a new wing pattern. The correction layer absorbs dynamics knowledge from world models and applies small, targeted corrections to VLA outputs — preserving everything the base model learned while adding domain-specific expertise.

ANIMA MODULE DEPENDENCIES

PYGMALION
Provides the frozen VLA backbone and latent representations for alignment

DEPLOYMENT STACK

  • Correction layer training pipeline with world model dynamics priors
  • Two-stage alignment: latent bridging + residual learning
  • Lightweight inference addon — minimal latency overhead on base VLA
  • Correction layer versioning and A/B testing infrastructure