Skip to content
5ColorPublic

About

Replay-free continual visuomotor control with hyperspherical routing and shared motor bases

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

DTRA: Dynamic Topology Ring Architecture

English | 한국어

Preprint (DTRA v3.0) arXiv Manuscript Zenodo v3.0.0 DOI Zenodo Concept DOI Zenodo ABK-VLA License: MIT Python 3.10+ PyTorch 2.x

Zero raw-demonstration replay, Task-ID-free continual visuomotor control with hyperspherical routing and shared motor bases (retaining lightweight $\le 96$-vector key-validation exemplars/task during registration): 87.00 ± 1.06% final LIBERO-100 success across three training seeds (ddof=1, verified at 85.20 ± 0.43% over 12,000 zero-retrain rollouts at n_eval=20, outperforming matched-encoder DER++, L2P, and O-LoRA).

Repository Separation Note: This repository tracks the full ARIN $\to$ Drawing-VLA $\to$ ABK-VLA $\to$ DTRA v3.0–v6.0 research lineage across two generations:

  1. Gen 1 — v1_drawing_vla (Drawing-VLA v1.0.0 & ABK-VLA v1.1.3): Closed-loop ACT sketch chunking ($K=16$, Catmull-Rom brush rendering in Drawing-VLA, DOI 10.5281/zenodo.24122154) followed by clamped cubic B-Spline control-knot compression ($N_{\text{cp}}=6 \to H=16$), geometric DoG prototype routing, and continual 2D/3D benchmarks (ABK-VLA, DOI 10.5281/zenodo.24212130).
  2. Gen 2 — dtra (DTRA v3.0 — Dynamic Topology Ring Architecture, corresponding to 5Color/dtra): Resolves Voronoi encroachment, duplicate-instruction routing collapse, and $O(K)$ task-specific head growth across $K=100$ sequential 7-DoF LIBERO tasks via $\mathbb{S}^{63}$ topological ring isolation, Two-Stage Top-$M$ CMC, SharedBasisBank ($100 \to 27$ frozen primitive bases, -73%), and Two-Timescale Snap-Back (post_snap_norm = 0.0000).
  3. Post-v3 Embodied Continual Tracks (v4.0–v6.0): Strictly excluded from DTRA v3.0 preprint tables and isolated under dtra/experiments/{robocasa_v4,calvin_v4,calvin_v5}/ and dtra/outputs/calvin_v6/ (archived alongside outputs/{organ_v4,calvin_v4,calvin_v5,calvin_v6,robocasa}/). Note that v6.0's OrganSystem2Planner borrows the Organ name from ARIN as a lightweight System 2 routing prototype, distinct from the conceptual multi-ring $10^4\text{--}10^8$ Organ roadmap.

Inference-input scope: Task-ID-free means no integer task index is passed to the policy at inference. Routing uses dual-camera features (plus depth in pure RGB-D tracks), proprioception, and a language embedding; evaluations that provide language commands are command-conditioned, not observation-only.

For the complete hypothesis-by-hypothesis chronology, see RESEARCH_LINEAGE.md, RELEASE_NOTES_v3.1.0.md (prior: v3.0.0), dtra/README.md, and dtra/PAPER_20P_OUTLINE.md.


Hero Rollout & Continual Scaling (LIBERO-100, $K=100$)

DTRA Elastic — Live Closed-Loop Rollouts from Saved $K=100$ Checkpoint (Replay: 0%) 7-Arm Strict-Parity Scaling Benchmark (K=10 -> 100, Full N=3 + DER++/L2P/O-LoRA)
DTRA LIBERO-100 K=100 Hero Rollout LIBERO-100 7-Arm Scaling Curves
4 Verified 100%-Success Clips from One Frozen $K=100$ Policy (Seed 42): T09 (libero_10 Kitchen Microwave — retained after 90 more tasks), T05 vs. T87 (Dup-Lang Pair — identical prompt routed cleanly to #05 vs. #87 while reusing Basis #04), and T67 (Living Room). 87.00 ± 1.06% ($n=5$, $N=3$, ddof=1) | 85.20 ± 0.43% over 12,000 zero-retrain rollouts ($n=20$, unseen init[5:20]: 84.80 ± 0.44%). Under identical frozen ResNet-18 parity, DTRA Elastic (87.00%) & DTRA Rigid (86.67%) compress 100 skills into 27 frozen bases (-73% basis slots / -5.33% total params) with 99.68% Dup-Lang routing, outperforming Frozen MLP (70.33%), ER 5% (39.47%), DER++ (38.40%), L2P (1.20%), and O-LoRA (1.00%).

1. Repository Structure

drawing-vla/
├── README.md                              # Project landing page, 3-seed v3.0/v3.1 tables & reproduction guide
├── RELEASE_NOTES_v3.1.0.md                # Official DTRA v3.1.0 arXiv final release notes (23p, ddof=1)
├── RELEASE_NOTES_v3.0.0.md                # Prior DTRA v3.0.0 release notes & Zenodo DOI (10.5281/zenodo.24297711)
├── RESEARCH_LINEAGE.md                    # Gen-1 -> Gen-2 research chronology & milestone ledger
├── requirements.txt                       # Core Python dependencies
│
├── v1_drawing_vla/                        # [Gen 1 Archive] B-Spline Action Chunking & ABK-VLA
│   ├── src/                               # Early B-Spline policies, evaluators & CL experiments
│   ├── handvsai/                          # HandVsAI crowdsourced scoring & sketch templates
│   └── paper_v1/                          # ABK-VLA preprint PDF & v1 architecture specs
│
├── dtra/                                  # [Gen 2 Main Research] Dynamic Topology Ring Architecture
│   ├── assets/                            # Lightweight README preview GIFs (< 8 MB, 640x360 @ 12 fps)
│   ├── PAPER_20P_OUTLINE.md               # Initial 20-page NeurIPS/CoRL 1-col outline & v3.0/v4.0 boundary spec
│   ├── core/
│   │   ├── bspline_core.py                # Frozen clamped C^2 cubic B-Spline basis decoder (N_cp=6 [2D/Meta-World] or N_cp=8 [7-DoF LIBERO] -> H=16)
│   │   ├── hypertube_transport.py         # [v4.0+ Track] Object-centric C^2 hypertube transport adapter
│   │   ├── sim_dtra_tube_dynamics.py      # HypertubeRingLayer, TwoStageDTRA & 6 invariant checks
│   │   └── topological_bridge.py          # C^2 geodesic bridge & myelinated transition graph core
│   ├── experiments/
│   │   ├── libero/                        # [v3.0] LIBERO-Goal (K=5) & LIBERO-100 (K=100) harness
│   │   ├── metaworld/                     # [v3.0] 8-task 3D Meta-World 3-seed 6-arm ablation & OOD suite
│   │   ├── calvin_v4/, calvin_v5/         # [Post-v3 Track] CALVIN 34-task System 1 & Z-gating anchors
│   │   └── robocasa_v4/                   # [Post-v3 Track] RoboCasa 12-task pure RGB-D kitchen CL
│   ├── outputs/
│   │   └── calvin_v6/                     # [Post-v3 Track] CALVIN v6 System 1+2 JSONs, HUD & scope notes
│   ├── visualization/
│   │   └── visualize_dtra_xai.py          # Generates 300 DPI architecture, OOD, manifold & telemetry figures
│   ├── README.md                          # DTRA package overview (English)
│   ├── README_KO.md                       # DTRA package overview (Korean)
│   └── paper_v2/
│       ├── DTRA_v3.0_Preprint.pdf         # Final 23-page DTRA preprint (XeTeX build, >=8pt tables)
│       ├── dtra_preprint.pdf              # Final 23-page clean ASCII arXiv build (pdfLaTeX/XeTeX safe)
│       ├── dtra_arxiv_submission.tar.gz   # Standalone arXiv source submission bundle (verified build)
│       ├── dtra_v3.0_preprint.tex         # XeTeX 23-page LaTeX source
│       ├── dtra_preprint.tex              # Pure ASCII 23-page arXiv LaTeX source
│
├── outputs/
│   ├── libero/                            # Locked LIBERO-100 3-seed JSONs (42, 100, 2026) & LIBERO-Goal pilot
│   ├── metaworld/                         # Meta-World 3-seed & supplemental JSON results + checkpoints
│   ├── figures/                           # 300 DPI publication figures (fig0 .. fig8)
│   ├── calvin_v4/, calvin_v5/, calvin_v6/, organ_v4/, robocasa/  # [Post-v3 Track] Isolated A6000 archives
│   └── v1_archive/                        # Archived Gen-1 benchmark plots & JSON logs
│
├── scripts/
│   ├── build_dtra_v3_paper.py             # Generates 23-page LaTeX + compiles both PDFs
│   ├── generate_libero100_figures.py      # Generates 300 DPI fig6, fig7 (3-seed ddof=1 error bars) & fig8
│   ├── summarize_libero_multiseed.py      # Prints finalized K=10..100 multi-seed Mean ± Sample Std (ddof=1) tables
│   └── sync_remote_dtra.py, check_remote_env.py
│
└── tests/                                 # Automated invariance & v3.0 module verification suite

2. Verified DTRA v3.0 / v3.1 Benchmark Results (LIBERO-100 & Meta-World Only)

A. LIBERO-100 (libero_10 + libero_90, $K=10 \to 100$) Multi-Seed & External CL Summary (Strict Encoder Parity)

All parity arms share identical frozen perception (ResNet-18 L1–L3 DualCam + proprioception) and LightweightTextEncoder with DTRA Elastic peak VRAM of 2,713 ± 94 MB (n_eval=5 per task at each milestone, Mean ± Sample Std ddof=1 across all tables). All four core arms cover complete N=3 training seeds {42, 100, 2026}, complemented by matched-encoder published continual learning baselines (DER++, L2P, O-LoRA) on Seed 42:

Arm (Strict Parity) $K=10$ SR $K=30$ SR $K=50$ SR $K=70$ SR $K=100$ SR $\uparrow$ FR $\downarrow$ BWT $\uparrow$ RouteAcc $\uparrow$ Dup-Lang Route $\uparrow$ Bases ($K=100$) Active Latency
arm_frozen_mlp ($N=3$) 82.00±5.29% 85.11±3.42% 79.07±0.61% 76.48±1.19% 70.33 ± 1.45% 2.90±1.28% +0.20±2.81% 81.94±0.29% 41.76 ± 2.56% 100 6.46±0.70 ms
arm_er_5pct ($N=3$) 58.67±6.11% 56.67±8.19% 52.00±5.41% 47.43±4.98% 39.47 ± 1.81% 26.33±4.24% -20.00±3.17% --- --- 1 4.93±0.72 ms
arm_dtra_rigid ($N=3$) 83.33±2.31% 90.44±1.02% 90.67±1.15% 87.52±0.87% 86.67 ± 0.81% 2.56±1.11% +0.67±1.35% 99.91±0.16% 99.68 ± 0.55% 27 8.70±1.14 ms
arm_dtra_elastic ($N=3$, Full) 80.00±2.00% 91.11±1.68% 90.67±1.22% 87.71±0.86% 87.00 ± 1.06% 2.69±0.82% +0.20±0.88% 99.90±0.16% 99.68 ± 0.55% 27 9.55±0.24 ms
DER++ (5% Buf. + Distill., Seed 42) 68.00% 60.00% 60.40% 51.71% 38.40% 29.70% -16.16% --- --- 1 5.49 ms
L2P (Prompt Pool, Seed 42) 8.00% 2.00% 1.60% 1.14% 1.20% 51.72% -51.72% --- --- 32 7.79 ms
O-LoRA (Ortho Subspace, Seed 42) 2.00% 1.33% 2.00% 1.14% 1.00% 42.63% -42.63% --- --- 100 7.05 ms
  • Per-Seed K=100 Breakdown (N=3 Complete, Sample Std ddof=1):
    • arm_dtra_elastic: Seed 42 87.80%, Seed 100 87.40%, Seed 2026 85.80% $\rightarrow$ Mean 87.00 ± 1.06% (+16.67 pp SR and +57.93 pp Dup-Lang routing over arm_frozen_mlp, 27 frozen shared bases = 73% basis-slot / 5.33% total parameter reduction).
    • arm_dtra_rigid: Seed 42 86.20%, Seed 100 86.20%, Seed 2026 87.60% $\rightarrow$ Mean 86.67 ± 0.81%.
    • arm_frozen_mlp: Seed 42 72.00%, Seed 100 69.60%, Seed 2026 69.40% $\rightarrow$ Mean 70.33 ± 1.45%.
    • arm_er_5pct: Seed 42 40.20%, Seed 100 40.80%, Seed 2026 37.40% $\rightarrow$ Mean 39.47 ± 1.81%.

Zero-Retrain Extended Initial-State Verification (n_eval=20, 12,000 Total Rollouts)

Evaluating all three saved $K=100$ checkpoints (Seeds 42, 100, 2026) across 20 initial states/task (6,000 rollouts/arm, 12,000 total; eval_libero100_extended_neval20.json; Mean ± Sample Std ddof=1) confirms zero overfitting to init[0:5]:

Saved $K=100$ Policy ($N=3$) init[0:5] Unseen Held-Out init[5:20] [95% Wilson CI] All init[0:20] [95% Wilson CI] Dup-Lang SR (Route) Unique-Lang SR
arm_frozen_mlp (6,000 rollouts) 70.00 ± 0.92% 68.82 ± 0.91% [67.5, 70.2] 69.12 ± 0.90% [67.9, 70.3] 36.33% (41.80%) 83.17%
arm_dtra_elastic (6,000 rollouts) 86.40 ± 1.06% 84.80 ± 0.44% [83.7, 85.8] 85.20 ± 0.43% [84.3, 86.1] 89.06% (99.94%) 83.55%
Gain ($\Delta$ DTRA $-$ MLP) +16.40 pp +15.98 pp +16.08 pp +52.72 pp (+58.14 pp) +0.38 pp

B. 3D Meta-World (8 Tasks, 3-Seed Causal Ablation & External CL, ddof=1) & LIBERO-Goal (5 Tasks)

Benchmark Arm / Mode Clean / Mean SR FR $\downarrow$ RouteAcc Severe OOD ($\sigma=0.20$) / Peak VRAM Source
Meta-World (8 Tasks, 3 Seeds) arm4_frozen_mlp_multimodal 85.6 ± 3.8% 0.0 ± 0.0% 100.0% 23.6 ± 4.8% OOD SR ablate_dtra_scientific_results.json
Meta-World (8 Tasks, 3 Seeds) arm6_dtra_full_multimodal (Feat-Reg) 72.2 ± 8.3% 0.0 ± 0.0% 100.0% 51.4 ± 5.4% OOD SR (1.87 ms $O(1)$ CPU) ablate_dtra_scientific_results.json
Meta-World Causal (3 Seeds) DTRA Full (No Feat-Reg, p=0, σ=0) 83.1 ± 6.3% 0.0 ± 0.0% 100.0% 45.0 ± 6.6% OOD SR ablate_dtra_v31_final_results.json
Meta-World Causal (3 Seeds) Frozen MLP + Conv4 Feat-Reg 90.8 ± 3.6% 0.0 ± 0.0% 100.0% 49.2 ± 10.1% OOD SR ablate_dtra_v31_final_results.json
Meta-World External CL (3 Seeds) DER++ / L2P / O-LoRA 73.3% / 16.7% / 7.2% 22.9% / 79.4% / 63.2% --- 13.6% / 17.5% / 9.4% OOD SR ablate_dtra_v31_final_results.json
Meta-World Reverse ($T_7 \to T_0$, Seed 42) arm6_dtra_full_multimodal 74.2% 0.0% 100.0% 64.2% OOD SR ablate_dtra_supplemental_results.json
LIBERO-Goal Pilot (5 Tasks, Seed 42) arm_dtra_elastic 92.0% 5.0% 100.0% 3,696 MB VRAM (post_snap_norm = 0.0) libero_goal_pilot_4arms.json

Evaluation scope and resource interpretation

All four core LIBERO-100 arms (Frozen MLP, ER 5%, DTRA Rigid, DTRA Elastic) cover three training seeds (42, 100, 2026, unified sample std ddof=1); mid-stream checkpoints use n_eval=5 rollouts/task, while final $K=100$ generalization for DTRA Elastic vs. Frozen MLP is verified across all three seeds over n_eval=20 initial states (12,000 rollouts total, including 15 unseen held-out initial states init[5:20] with descriptive pooled 95% Wilson CIs). Matched-encoder external baselines (DER++, L2P, O-LoRA) evaluate Seed 42 on LIBERO-100 and all three seeds on Meta-World.

Parameter freezing ($\nabla_\theta=0$) and temporary state reset are separate: on LIBERO-100, episode_snap_back() clears routing inertia and write-only diagnostic buffers (post_snap_norm = 0.0000, so chunk_snap_back() does not alter forward control outputs); active per-chunk snap_back() applies to Meta-World. LIBERO-100 empirical success-rate forgetting is 2.69 ± 0.82% (DTRA Elastic) and 2.56 ± 1.11% (DTRA Rigid).

The 73% reduction concerns motor basis slot count (100 -> 27 frozen bases, 3.81 MB vs. 14.12 MB unshared). Because per-task LIBEROHypertubeNode adapters occupy 179.33 MB (97.92% of total policy memory), SharedBasisBank yields a 5.33% total policy-parameter reduction (193.45 -> 183.14 MB, vs. 151.67 MB for Frozen MLP and 1.70–2.09 MB for monolithic baselines, excluding the common frozen backbone and non-parameter buffers).

On Meta-World, our three-seed causal ablation (ablate_dtra_v31_final.py, ddof=1) isolates the clean-versus-OOD trade-off: removing Conv4 feature-noise regularization (p=0, σ=0) lifts DTRA Full from 72.2 ± 8.3% to 83.1 ± 6.3% clean SR (+10.8 pp unrounded) while preserving 45.0 ± 6.6% severe OOD SR (vs. unregularized Frozen MLP's 23.6 ± 4.8%). Broad retrieval remains $O(K)$, with a constant-cost selected decoding core (9.55 ± 0.24 ms GPU policy inference at $K=100$).

C. Post-v3 Embodied Continual Tracks (DTRA v4.0 – v6.0, Separate from Preprint)

Track / Benchmark Arm / Mode Key Metrics Drift Source
v4.0 RoboCasa (12 Tasks: 8 Atomic + 4 Composite, 0% Replay) Arm4_DTRA_v4_Hypertube (vs Flat_MLP/ER_5% 0.00%) Atomic SR 29.17% (OpenMicrowave 100.0%), Overall 19.44%, RouteAcc 100.0%, BWT +0.00%, 1.928 ms (Composite 0.00% $\rightarrow$ System 2 motivation) 0.000000 robocasa_continual_results_4arms.json
v5.0 + v6.0 CALVIN (34 Tasks, 30 5-Step Chains) Pure Vision (Leak: ZERO) RouteAcc 97.65%, Single SR 55.88%, 5-Chain Len 2.37 (vs 1.90), 5/5 SR 16.67% (vs 6.67%), Jerk +15.12%, 22 edges, 9.30 ms 0.0000 calvin_v6_vision_34tasks.json
v5.0 + v6.0 CALVIN (34 Tasks, 30 5-Step Chains) Oracle Anchor (Upper-Bound Ref) RouteAcc 97.65%, Single SR 46.47%, 5-Chain Len 2.53 (vs 1.90), 5/5 SR 16.67% (vs 10.00%), Jerk +15.24%, 25 edges, 8.49 ms 0.0000 calvin_v6_oracle_34tasks.json
Expand Post-v3 Embodied Rollout Previews (CALVIN v6 Pure Vision & RoboCasa v4 Pure RGB-D)
DTRA v6.0 — CALVIN Pure Vision 5-Step Chain (System 1 + System 2) DTRA v4.0 — RoboCasa Pure RGB-D Kitchen (Hypertube Transport)
DTRA v6.0 CALVIN Pure Vision 5-Step Chain DTRA v4.0 RoboCasa Pure RGB-D Rollout
Canonical 5/5 sequence (Chain #1, Mode: VISION, Leak: ZERO, Drift: 0.0000, live System 2 recovery CELL #108 -> #100). (Scope Notes) 12 tasks (8 Atomic + 4 Composite, 0% replay, no object-state, Drift: 0.000000). (Results JSON)

3. Reproduction and Code Availability

Core DTRA modules, unit tests, figure/paper build scripts, locked result JSONs, and the Meta-World / LIBERO / post-v3 benchmark scripts are included in this repository. Installing requirements.txt covers the core modules, unit tests, and figure/LaTeX regeneration scripts; full simulator rollouts additionally require installing the external simulation packages (metaworld, mujoco, gymnasium for Meta-World; libero, robosuite, bddl, h5py, and LIBERO HDF5 demos for LIBERO-100). Use python scripts/check_remote_env.py to inspect which simulator dependencies are available in your environment.

# 0. Install core dependencies, inspect environment & run invariant unit tests
pip install -r requirements.txt
python scripts/check_remote_env.py --env dtra
python -m unittest discover -s tests -v
python dtra/tests/test_topological_bridge.py

# 1. Step-1 Hypertube manifold invariants (6 automated checks)
python dtra/core/sim_dtra_tube_dynamics.py

# 2. 3D Meta-World 8-task 3-seed scientific ablation & supplemental stress suite (requires metaworld/mujoco)
python -m dtra.experiments.metaworld.ablate_dtra_scientific --seeds 42 100 2026 --epochs 55
python -m dtra.experiments.metaworld.ablate_dtra_supplemental --seed 42 --epochs 55

# 3. LIBERO-100 (K=100) continual scaling stream (requires `libero` conda env + LIBERO HDF5 demos)
python -m dtra.experiments.libero.benchmark_dtra_libero \
  --suites libero_10,libero_90 --max-tasks 100 \
  --stream-tasks --eval-milestones 10,30,50,70,100 \
  --epochs 25 --batch-size 128 --n-eval 5 --seed 42 \
  --out-json outputs/libero/benchmark_dtra_libero_10_libero_90_k100.json

# 4. Aggregate 3-seed tables, regenerate 300 DPI figures & compile 23-page DTRA PDFs
python scripts/summarize_libero_multiseed.py
python scripts/generate_libero100_figures.py
python scripts/build_dtra_v3_paper.py

4. Visual Figures (outputs/figures/) & Preprints (dtra/paper_v2/)


5. Citation & License

@misc{son2026dtra,
  author       = {Son, Seungjoon},
  title        = {{DTRA}: Dynamic Topology Ring Architecture for Replay-Free Continual Visuomotor Control via Hyperspherical Isolation and Shared Motor Bases},
  year         = {2026},
  month        = nov,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.24297711},
  url          = {https://doi.org/10.5281/zenodo.24297711},
  note         = {v3.1.0 23-page Preprint (v3.0.0 Zenodo archive); Concept DOI: 10.5281/zenodo.24218482; prior v2.1.0: 10.5281/zenodo.24218483; Repository: \url{https://github.com/5Color/dtra}}
}

@misc{son2026abkvla,
  author       = {Son, Seungjoon},
  title        = {{ABK-VLA}: Task-ID-Free Continual Visuomotor Policy Learning via Autonomous B-Spline Knot Compression and Visual Prototype Routing},
  year         = {2026},
  month        = nov,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.24212130},
  url          = {https://doi.org/10.5281/zenodo.24212130},
  note         = {v1.1.3; v1.1.1: 10.5281/zenodo.24208545; Concept DOI: 10.5281/zenodo.24208544}
}

Released under the MIT License.

About

Replay-free continual visuomotor control with hyperspherical routing and shared motor bases

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages