LightAcotVLA

LightAcotVLA is a research codebase focused on making action-space reasoning faster, more selective, and easier to verify.

The current policy combines three components:

• Explicit Action Reasoner (EAR): produces a 15-frame coarse action plan.
• Implicit Action Reasoner (IAR): extracts latent action priors from the shared vision-language representation.
• Final action expert: consumes EAR and IAR features and generates a 10-action control chunk.

The project studies where computation can be removed without hiding failures: execution-horizon selection, Action-CoT denoising reduction, failure-focused adaptation, and one-step EAR/final distillation.

flowchart LR
    O["Observation + instruction"] --> V["Vision-language prefix"]
    V --> I["Implicit action reasoner"]
    V --> E["Explicit Action-CoT reasoner"]
    I --> F["Final action expert"]
    E --> F
    F --> A["10-step action chunk"]
    A --> H["Execution-horizon controller"]

Loading

Status snapshot: 2026-07-24. The repository contains formal evaluations, diagnostic oracles, pilot training code, and the corresponding structured results. Not every pilot is a deployable improvement; evidence boundaries are stated explicitly below.

Current progress

1. Execution Horizon V2-P

The execution-horizon line controls how many environment steps are executed before the next expensive policy call. The formal comparison uses LIBERO-10, 10 tasks × 100 trials, seed 7, paired initial states, 10 Action-CoT denoising steps, and real policy/server/RPC timing.

Scheme	Success	Calls / episode	Avg. H	Policy s / episode	Interpretation
Paired H5 baseline	92.7%	63.518	5.000	5.687	Seeded project baseline
Fixed H9	95.0%	33.568	9.000	2.983	Strongest simple efficiency reference
Exact batched-MC K20	94.4%	34.647	8.846	4.230	Analysis teacher; not deployable
V2-P distilled	95.2%	30.536	9.874	3.387	Fewer calls, but predictor overhead remains
V2-P value-refined	94.3%	32.892	9.328	3.611	No Pareto improvement over Fixed H9

The distilled selector mostly chooses H10. Its success point estimate is slightly above Fixed H9, but its policy time per episode is higher. The current evidence therefore does not establish a dynamic-selector Pareto win. This table reports the project's seeded 100-trial-per-task protocol and should not be compared directly with evaluations that use only 50 trials per task.

Evidence:

• formal human-readable report
• formal JSON/CSV outputs
• headroom, hard-state, RL, candidate, and progress diagnostics

2. Action-CoT compute reduction

Stage B separates two ideas that are easy to conflate:

1. changing or masking coarse Action-CoT values;
2. actually reducing the number of coarse flow denoising steps.

Only the second directly removes model computation in the current dense architecture.

Coarse denoising steps	Open-loop full ACoT latency	Reduction vs. 10 steps
10	73.82 ms	0.00%
7	67.42 ms	8.67%
5	63.23 ms	14.35%
3	60.32 ms	18.29%
1	57.10 ms	22.65%

These are open-loop timing results. Token pruning or cached overrides that still execute the same dense expert do not provide the same compute saving, and closed-loop quality must be evaluated separately. The table varies coarse NFE while final-action denoising remains fixed at 10 steps.

Evidence: Stage B results.

3. Task8/9 failure-focused adaptation

Task8 and Task9 are the main failure-focused testbed. The repository includes targeted SFT, a uniform continued-SFT control, hard-state DAgger collection, and held-out Fixed H9 evaluation.

Checkpoint / protocol	Episodes	Success
Targeted SFT 1k, initial 20 roots per task	40	80.0%
Uniform control 1k, same initial roots	40	95.0%
Uniform control 1k, episodes 20–99	160	91.25%
Base 50999, 25 unseen roots × 2 policy-seed rounds/task	100	89.0%
DAgger 1k, same roots and policy-seed rounds	100	91.0%

The targeted SFT pilot regressed on its paired initial set relative to the uniform control. Round-1 hard-state collection accepted only 3 corrective trajectories (705 frames, 2 roots) from 33 attempted hard roots. The DAgger 1k checkpoint had 8 rescues and 6 regressions relative to base (McNemar p=0.7905): the two policy-seed halves were net -1 and +3. It therefore did not pass the preset continuation gate and is not a stable improvement.

Evidence: Task8/9 adaptation results.

4. IR-ACoT one-step pilot

IR-ACoT tests a different speedup axis: distill the original 10-step EAR and 10-step final flow into one-step students while keeping the VLM, IAR, and reasoning fusion frozen. IR-lite adds an intervention-response loss; B6-lite is the matched clean endpoint control with the same deployed graph.

Completed:

• a 200-state open-loop EAR intervention audit;
• 2,000 clean teacher endpoint exports;
• 300-step one-step EAR training;
• 2,000 one-step student-EAR exports;
• 2,000 teacher-final relabels conditioned on those student EARs;
• 300-step B6-lite and IR-lite final-student training.

The causal audit passed because the final action responded strongly to EAR rotation and partly to translation. This is an open-loop sensitivity result, not a success-rate result.

Final 300-step checkpoint	Validation final MSE	Validation IR-delta MSE	IR-delta cosine
B6-lite	0.003643	0.000309	0.502
IR-lite	0.003630	0.000311	0.496

At the final checkpoint, IR-lite has not demonstrated a response-alignment advantage over the matched B6-lite control. Closed-loop LIBERO success and real one-step staged latency have not yet been evaluated, so no method-quality or end-to-end speedup claim is made.

Evidence:

• pilot protocol
• pilot summaries and metric streams

Evidence map

Path	Purpose
reports/context/experiment_log.md	Append-only factual project ledger; about 480 KB at this snapshot
reports/README.md	Report and result index
results/execution_horizon_v2p	Formal V2-P results and diagnostics
results/stage_b	Action-CoT pruning, denoising, timing, and closed-loop summaries
results/task89_adaptation	Targeted SFT, uniform control, and DAgger data
results/ir_acot_pilot	One-step EAR/final training and audit data

Presentation files, checkpoints, HDF5 shards, videos, raw per-step rollouts, full training logs, environments, and caches are intentionally not tracked.

Main implementation areas

Area	Main files
Core policy model and profiling	src/openpi/models/acot_vla.py, src/openpi/policies/policy.py
Execution-horizon predictor	src/openpi/models/execution_horizon_predictor.py, src/openpi/execution_horizon/
V2-P evaluation and training	scripts/eval_libero_execution_horizon.py, scripts/train_execution_horizon_predictor.py
Headroom and selector diagnostics	scripts/audit_execution_horizon_headroom.py, scripts/audit_execution_horizon_selector_eval.py
Stage B pruning and NFE sweeps	scripts/eval_libero_action_cot_pruning.py, scripts/sweep_action_cot_denoising_steps.py
Hard-state DAgger	scripts/collect_libero_hard_state_dagger.py
IR-ACoT endpoint distillation	src/openpi/action_cot/endpoint_dataset.py, scripts/audit_ear_interventions.py, scripts/train_acot_endpoint_distillation.py

New evaluation and sidecar paths are optional. Ordinary training, serving, and evaluation behavior remains unchanged unless the corresponding flags are used.

Getting started

Installation

The project uses uv for dependency management.

git clone https://github.com/KongyueX/LightAcotVLA.git
cd LightAcotVLA
git submodule update --init --recursive
GIT_LFS_SKIP_SMUDGE=1 uv sync
GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .

Base training and serving

# Convert LIBERO data to LeRobot format.
python examples/libero/convert_libero_data_to_lerobot.py

# Compute normalization statistics.
uv run scripts/compute_norm_stats.py --config-name <CONFIG_NAME>

# Train and serve.
bash scripts/train.sh <CONFIG_NAME> <EXP_NAME>
bash scripts/server.sh <GPU_ID> <PORT>

Experiment entry points

uv run python scripts/eval_libero_execution_horizon.py --help
uv run python scripts/eval_libero_action_cot_pruning.py --help
uv run python scripts/collect_libero_hard_state_dagger.py --help
uv run python scripts/audit_ear_interventions.py --help
uv run python scripts/train_acot_endpoint_distillation.py --help

The IR-ACoT server workflow is documented in docs/ir_acot_mvp.md.

Next gates

1. Evaluate original 10×10, untrained 1×1, B6-lite 1×1, and IR-lite 1×1 under the same Fixed H9 closed-loop protocol and measure staged latency.
2. Continue IR-ACoT only if it improves the success/latency Pareto point over B6-lite; open-loop reconstruction alone is insufficient.
3. Do not repeat the same global DAgger recipe unchanged; first improve correction coverage or add state-selective recovery, then compare across multiple policy seeds.
4. Keep Fixed H9 as the execution-horizon efficiency reference; add selector complexity only when held-out rescues exceed regressions at comparable policy time.

License and third-party notices

LightAcotVLA is distributed under the Apache License 2.0. The repository contains modified third-party open-source components; source provenance and license notices are recorded in THIRD_PARTY_NOTICES.md.