PACE: Predictability-Aware Closed-loop Execution for Long-Horizon Manipulation

Predictability cliffs are intrinsic structures of long-horizon manipulation tasks. This repository implements the cliff detection, closed-loop replanning, and phenomenon discovery experiments from the PACE v2 paper (CoRL 2027 submission).

Implementation Status

Read this table before any number in this README. Anything tagged placeholder is a synthetic dry-run output; only verified items are checked numerically against the math.

Item	Status	Notes
Phase InfoNCE identifiability ($\hat I^{(1)}$ infrastructure)	verified (CPU)	`verify_identifiability.py` PA $\ge 0.7$
$\hat I^{(1)}$ Bhattacharyya $\beta_t$	verified (CPU)	`verify_phase_posterior.py` peak-F1 $\ge 0.5$
PCAR budget-quantile DKW bound	verified (CPU)	`verify_pcar_budget.py` $\|\text{rate}-\epsilon\|<0.005$
Boundary-aware flow loss reduction	verified (CPU)	`sanity_pace_a.py` $\ge 20%$ FM drop
Unit + smoke tests	verified (CPU)	`pytest tests/` → 212 passed (v2.1)
$\hat I^{(2)}$ Action variance	verified (CPU)	`compute_I_hat_2` implemented; tests in `test_cliff_estimators.py`
$\hat I^{(3)}$ Velocity curvature	verified (CPU)	`compute_I_hat_3` implemented; cached `c_{t-1}` cleared by `policy.reset()`
Concordance $C_t$ rank fusion	verified (CPU)	`compute_concordance_C` rolling-window rank fusion; tests cover [0,1] range and online state
LIBERO-Long / Spatial SR (Tables 1, 2)	placeholder	requires GPU + LIBERO dataset + trained checkpoint
Phenomenon §6.1, §6.4 numbers	placeholder	dry-run pipeline; full results require real env + checkpoints
Inference cost (params, NFE, latency)	placeholder	parameter count and NFE are correct per architecture; latency requires real GPU benchmark

This environment is CPU-only with no LIBERO/SimplerEnv installed and no trained checkpoints. The CPU-verifiable items above are what this repo can prove correct today; everything tagged placeholder needs a GPU + datasets to produce real numbers (see Verification & Reproduction Guide below).

Quick Start

# 1. Install (CPU is enough for verification scripts and unit tests)
conda env create -f environment.yml && conda activate lerobot_env
pip install -r requirements.txt
pip install -e ./lerobot_policy_phaseqflow

# 2. Smoke test (no GPU, no data required)
bash scripts/smoke/smoke_phase_centric.sh

# 3. Run all CPU-only verifications (~3 min total)
pytest tests/ -q
python scripts/verification/verify_identifiability.py
python scripts/verification/verify_phase_posterior.py
python scripts/verification/sanity_pace_a.py
python scripts/verification/verify_pcar_budget.py

See docs/OPERATIONS_GUIDE.md for the engineering handbook and docs/ARCHITECTURE.md for the full specification, including the math behind each verification script.

Main Results

Table 1 — LIBERO-Long success rate: published baselines

Status: GPU evaluation not yet run. Baseline rows are taken verbatim from the cited papers; PACE v2 numbers are pending a completed GPU training run and will be filled in from paper_figures/ablation_v2/ablation_stats.csv.

Method	LIBERO-Long SR	Source
OpenVLA-OFT	54.5%	arXiv 2502.19645, Table 2
π₀	60.0%	arXiv 2410.24164, Table 1
MoH	57.8%	arXiv 2410.11842, Table 3
PACE v2 config 07 (ours)	pending GPU run	—
Oracle cliff (config 06, upper bound)	pending GPU run	—

Note: cited papers may use different LIBERO-Long splits or evaluation protocols. Direct numeric comparison requires a shared evaluation harness.

Ablation Study (Table 2)

Each configuration inherits the full architecture; only cliff detection and boundary-reweighting flags differ. CoRL submission uses three configs × 3 seeds, evaluated with IQM ± 95% bootstrap CI (rliable-style):

Status: all numbers below are synthetic dry-run placeholders generated by scripts/aggregate_ablation.py --dry_run. Real numbers land after the cloud sweep finishes (scripts/run_autodl_pipeline.sh train).

v2.1 implementation status: compute_I_hat_1 (Bhattacharyya), compute_I_hat_2 (action variance), compute_I_hat_3 (velocity curvature), and compute_concordance_C (rank-window fusion) are all implemented and exercised by tests/test_cliff_estimators.py. The CoRL submission's headline ablation 07 uses all three estimators internally via the concordance signal, so the trimmed table still validates the full theoretical chain.

Config	Description	LIBERO-Long IQM (placeholder)	LIBERO-Spatial IQM (placeholder)
01	BC-Chunked (baseline)	0.520	0.634
02	Cliff via β̂_t only (I^(1))	0.593	0.690
07	PACE v2: C_t + boundary reweight	0.692	0.727

All numbers above are synthetic dry-run placeholders pending the full LIBERO-Long / LIBERO-Spatial GPU sweep.

# Verify pipeline (synthetic data, no checkpoint):
python scripts/aggregate_ablation.py --dry_run

# Real aggregation after GPU training:
python scripts/aggregate_ablation.py \
    --input_root outputs/ablation_v2 \
    --output paper_figures/ablation_v2/

Phenomenon Results (Section 6)

All numbers in §6.1–6.5 are placeholder outputs from --dry_run mode (synthetic distributions hard-coded inside each script). They verify the analysis pipeline but do not reflect real LIBERO rollouts. Real numbers are produced by running each script without --dry_run against trained checkpoints; see the Verification & Reproduction Guide.

§6.1 Universality — cliff occurrence across 4 policies (placeholder)

The pipeline rolls out OpenVLA-7B, π₀, BC-ACT, and Diffusion Policy on LIBERO-Long, runs cliff detection, and measures the time gap between the last detected cliff and the first failure step. The hypothesis is that the gap distribution is policy-agnostic (right-skewed, mass within 5–25 steps of the last cliff). Pairwise KS test $p < 0.05$ would confirm cliffs are not specific to any one policy.

# Placeholder (CPU, synthetic):
python scripts/phenomenon/universality.py --dry_run

# Real (requires the four checkpoints + LIBERO-Long environment):
python scripts/phenomenon/universality.py \
    --n_rollouts 50 --seeds 0 1 2 \
    --output paper_figures/universality/

§6.2 Regret Scaling — δSR vs. chunk length H (placeholder)

Hypothesis: success-rate regret $\delta\text{SR} = \text{SR}_{\text{ref}} - \text{SR}$ scales with chunk length $H$ as $\delta\text{SR} \approx c \cdot H \cdot \overline{\Delta H}$ (Proposition 3 in the paper).

Chunk length $H$	SR (placeholder)	$\delta$SR (placeholder)
4	0.829	0.000
8	0.846	0.000
16	0.712	0.007
32	0.814	0.025
64	0.718	0.048

Linear fit through origin reports $R^2 \approx 0.82$ on the placeholder synthetic data; real expected $R^2$ is reported only after a real LIBERO sweep.

# Placeholder (CPU, synthetic):
python scripts/phenomenon/regret_scaling.py --dry_run

# Real (requires PACE v2 checkpoint + LIBERO-Long):
python scripts/phenomenon/regret_scaling.py \
    --checkpoint checkpoints/pace_v2_libero_long \
    --H_values 4 8 16 32 64 \
    --output paper_figures/regret_scaling/

§6.3 Triangulation Concordance (placeholder)

Precision/recall of each estimator and the concordance $C_t$ on a gripper-flip oracle, $\pm 5$-step tolerance:

Detector	Precision	Recall	F1
$\hat I^{(1)}$ Bhattacharyya $\beta_t$	0.162	1.000	0.279
$\hat I^{(2)}$ Action variance $\sigma_t^2$	0.155	1.000	0.269
$\hat I^{(3)}$ Velocity curvature $\kappa_t$	0.153	1.000	0.266
Concordance $C_t$	1.000	1.000	1.000

All numbers above are placeholders. The trivial 1.000/1.000/1.000 row for $C_t$ comes from a synthetic generator where the three estimator spikes are perfectly correlated by construction; real numbers require $\hat I^{(2)}$ and $\hat I^{(3)}$ to be implemented (see ARCHITECTURE.md §3.3–3.5).

python scripts/phenomenon/triangulation_concordance.py --dry_run

§6.4 Alternative Cliff Detectors (placeholder)

Comparison of six detector strategies on a gripper-flip oracle, $\pm 5$-step tolerance:

Detector	Precision	Recall	F1
Concordance $C_t$ (PACE v2)	0.429	1.000	0.600
Bhattacharyya $\beta_t$	0.353	1.000	0.522
KL divergence $D_{\text{KL}}(p_t \| p_{t-1})$	0.347	1.000	0.515
JS divergence $D_{\text{JS}}$	0.353	1.000	0.522
Posterior entropy $H(\hat p)$	0.353	0.973	0.519
BOCPD	0.071	0.953	0.133

python scripts/diagnostics/diagnostic_utils/trigger_comparison.py --dry_run

§6.5 Boundary Loss Ratio (placeholder)

Hypothesis: the flow-matching loss at phase-boundary steps is materially higher than at interior steps, motivating boundary-aware reweighting. The placeholder generator hard-codes $\mathbb{E}{\text{boundary}} = 0.20$, $\mathbb{E}{\text{interior}} = 0.05$, producing a ratio of $\approx 4$. Real values come from sweeping boundary timesteps across LIBERO demos with a trained model.

python scripts/diagnostics/diagnostic_utils/measure_boundary_error.py --dry_run

Inference Cost (placeholder)

Parameter counts and NFE per inference path are derived from the architecture (correct by construction); the latency and Hz columns require a real GPU benchmark and are placeholders.

Method	Params	NFE	Latency (placeholder)	Hz (placeholder)
BC-Chunked	312.4M	4	12.4 ms	81
Cliff via $\beta_t$	312.5M	4	12.7 ms	79
Cliff via $\sigma_t^2$	312.5M	36	91.4 ms	11
Cliff via $\kappa_t$	312.5M	8	23.3 ms	43
Concordance $C_t$	312.6M	44	96.5 ms	10
PACE v2 ($C_t$ + reweight)	312.6M	44	99.1 ms	10

NFE derivation: $\beta_t$ adds 0 NFE (uses planner softmax, no extra flow forward); $\sigma_t^2$ requires $N=8$ samples × 4-NFE shortcut + 4 NFE base = 36; $\kappa_t$ uses two 4-NFE forwards = 8; concordance adds the three contributions minus shared base = 44.

# Placeholder latency (CPU; the latency column is hardcoded synthetic):
python scripts/diagnostics/diagnostic_utils/measure_inference_cost.py --dry_run

# Real (requires GPU + checkpoint):
python scripts/diagnostics/diagnostic_utils/measure_inference_cost.py \
    --checkpoint checkpoints/pace_v2_libero_long \
    --device cuda --batch_size 1

Verification & Reproduction Guide

This section is the complete recipe to take this repo from a clean checkout to the figures and tables in the paper. It is split into five parts in order of increasing resource demand:

Part A: CPU-only verifications of the math (no GPU, no data, $\le 5$ min)
Part B: Prerequisites for real experiments (GPU, datasets, checkpoints)
Part C: Three-stage training pipeline
Part D: Evaluation pipeline (LIBERO-Long, LIBERO-Spatial, SimplerEnv)
Part E: Phenomenon experiments and figure generation

Part A — CPU verification (no GPU, no data, ~5 min)

These verify mathematical claims on synthetic distributions and are the gate for any code change.

# A.1 Environment
conda env create -f environment.yml && conda activate lerobot_env
pip install -r requirements.txt
pip install -e ./lerobot_policy_phaseqflow

# A.2 Unit + smoke tests (~30 s)
pytest tests/ -q                              # expects: 204 passed
bash scripts/smoke/smoke_phase_centric.sh     # 7-mode E2E, all [OK]

# A.3 Math verifications (~3 min)
python scripts/verification/verify_identifiability.py    # PA >= 0.7 (or WARN_DEGENERATE on CPU)
python scripts/verification/verify_phase_posterior.py    # peak-F1 >= 0.5
python scripts/verification/sanity_pace_a.py             # boundary FM drop >= 20%
python scripts/verification/verify_pcar_budget.py        # |rate - eps| < 0.005

# A.4 Pipeline dry-runs (verifies orchestration; numbers are synthetic)
python scripts/aggregate_ablation.py --dry_run
python scripts/phenomenon/universality.py --dry_run
python scripts/phenomenon/regret_scaling.py --dry_run
python scripts/phenomenon/triangulation_concordance.py --dry_run
python scripts/diagnostics/diagnostic_utils/trigger_comparison.py --dry_run
python scripts/diagnostics/diagnostic_utils/measure_boundary_error.py --dry_run
python scripts/diagnostics/diagnostic_utils/measure_inference_cost.py --dry_run

If every line in A.2–A.3 returns PASS / [OK] / 204 passed, the implemented math is correct. Part A does not produce paper numbers.

Part B — Prerequisites for real experiments

Resource	Requirement	Where to get it
GPU	$\ge 24$ GB VRAM (RTX 5070 / A100 / H100)	local or cloud
LIBERO datasets	LIBERO-10 (long-horizon) and LIBERO-Spatial	https://libero-project.github.io
SimplerEnv	for §6 universality on RT-1 / OpenVLA simulator	https://github.com/simpler-env/SimplerEnv
Baseline checkpoints	OpenVLA-7B, π₀, BC-ACT, Diffusion Policy	HuggingFace; loaded via `baselines/*_adapter.py`
Storage	$\sim 200$ GB for datasets + checkpoints	—

The four baseline adapters in baselines/ already handle download + load via is_available() / load() / rollout(); populate HF_HOME and the adapter load() will fetch the weights on first call.

Once datasets are placed under data/libero/ and checkpoints under checkpoints/, the pipeline below runs without further manual fetch.

Part C — Three-stage training (real)

Stage 1–3 produce a single PACE v2 checkpoint. Each stage is a separate config under configs/train/.

# Stage 1: pretrain multimodal tokenizer (~2 days on RTX 5070)
python scripts/training/train_dummy_batch.py \
    --config configs/train/01_pretrain_multimodal.yaml \
    --total_steps 50000 --device cuda \
    --output checkpoints/stage1_pretrain/

# Stage 2: hierarchical phase encoder + boundary-aware flow head (~3 days)
python scripts/training/train_dummy_batch.py \
    --config configs/train/02_train_phase_and_flow.yaml \
    --resume checkpoints/stage1_pretrain/last.pt \
    --total_steps 100000 --device cuda \
    --output checkpoints/stage2_phase_flow/

# Stage 3: freeze main net, calibrate PCAR / B-PCAR / concordance (~6 hours)
python scripts/training/train_dummy_batch.py \
    --config configs/train/03_finetune_replan.yaml \
    --resume checkpoints/stage2_phase_flow/last.pt \
    --total_steps 5000 --device cuda \
    --output checkpoints/pace_v2_libero_long/

# Calibration helpers (run after stage 3):
python scripts/calibration/calibrate_concordance.py \
    --checkpoint checkpoints/pace_v2_libero_long \
    --output checkpoints/pace_v2_libero_long/concordance_threshold.json
python scripts/calibration/calibrate_b_pcar.py \
    --checkpoint checkpoints/pace_v2_libero_long \
    --output checkpoints/pace_v2_libero_long/b_pcar_params.json

Part D — Evaluation pipeline (real)

Evaluates the trained checkpoint on LIBERO-Long, LIBERO-Spatial, and SimplerEnv. Each rollout records SR + PCAR replan timings to eval_results.json.

# D.1 LIBERO-Long (50 episodes × 3 seeds)
bash scripts/eval/run_libero_main.sh \
    --checkpoint checkpoints/pace_v2_libero_long \
    --dataset LIBERO-10 \
    --seeds 42 123 2024 \
    --output outputs/eval/libero_long/

# D.2 LIBERO-Spatial (specificity test, §4.4 of ARCHITECTURE.md)
bash scripts/eval/run_libero_main.sh \
    --checkpoint checkpoints/pace_v2_libero_long \
    --dataset LIBERO-Spatial \
    --seeds 42 123 2024 \
    --output outputs/eval/libero_spatial/

# D.3 SimplerEnv (§6.1 universality)
bash scripts/eval/run_simpler.sh \
    --checkpoint checkpoints/pace_v2_libero_long \
    --output outputs/eval/simpler/

# D.4 LIBERO-Perturbed (robustness)
python scripts/eval/libero_perturbed.py \
    --checkpoint checkpoints/pace_v2_libero_long \
    --output outputs/eval/libero_perturbed/

Part E — Ablation, phenomenon, and figures (real)

After Part C produces a checkpoint and Part D produces evaluation JSONs, the ablation matrix and the phenomenon experiments fold both into the paper figures.

# E.1 Seven-config ablation matrix (21 GPU runs total)
for config in configs/ablation/v2/0[1-7]_*.yaml; do
  for seed in 42 123 2024; do
    python scripts/training/train_dummy_batch.py \
        --config $config --seed $seed --device cuda \
        --total_steps 20000 \
        --output outputs/ablation_v2/$(basename $config .yaml)/seed_$seed/
  done
done

# E.2 Aggregate ablation results (rliable IQM + bootstrap CI + Wilcoxon)
python scripts/aggregate_ablation.py \
    --input_root outputs/ablation_v2/ \
    --output paper_figures/ablation_v2/
# Produces: ablation_stats.csv, ablation_table_v2.tex (paper Table 2)

# E.3 Phenomenon experiments (§6 of paper)
python scripts/phenomenon/universality.py \
    --n_rollouts 50 --seeds 0 1 2 \
    --output paper_figures/universality/
python scripts/phenomenon/regret_scaling.py \
    --checkpoint checkpoints/pace_v2_libero_long \
    --H_values 4 8 16 32 64 \
    --output paper_figures/regret_scaling/
python scripts/phenomenon/triangulation_concordance.py \
    --checkpoint checkpoints/pace_v2_libero_long \
    --output paper_figures/triangulation/

# E.4 Diagnostic measurements (§6.4, §6.5, inference cost)
python scripts/diagnostics/diagnostic_utils/trigger_comparison.py \
    --checkpoint checkpoints/pace_v2_libero_long \
    --output paper_figures/diagnostics/
python scripts/diagnostics/diagnostic_utils/measure_boundary_error.py \
    --checkpoint checkpoints/pace_v2_libero_long \
    --output paper_figures/diagnostics/boundary_ratio.json
python scripts/diagnostics/diagnostic_utils/measure_inference_cost.py \
    --checkpoint checkpoints/pace_v2_libero_long \
    --device cuda --output paper_figures/diagnostics/inference_cost.json

# E.5 Generate paper figures (all read from paper_figures/*/ produced above)
python scripts/figures/fig1_universality.py \
    --input paper_figures/universality/raw_distances.json \
    --output paper_figures/fig1_universality.pdf
python scripts/figures/fig2_method_overview.py    # purely diagrammatic
python scripts/figures/fig3_phase_visualization.py \
    --input paper_figures/triangulation/phase_vis_data.json \
    --output paper_figures/fig3_phase_visualization.pdf
python scripts/figures/fig4_regret_scaling.py \
    --input paper_figures/regret_scaling/regret_vs_H.csv \
    --output paper_figures/fig4_regret_scaling.pdf
python scripts/figures/fig5_concordance_pr_curve.py \
    --input paper_figures/diagnostics/trigger_comparison.csv \
    --output paper_figures/fig5_concordance_pr_curve.pdf

# E.6 One-shot orchestrator (does C.E1 → E.5 with default settings)
bash scripts/run_experiments.sh --checkpoint checkpoints/pace_v2_libero_long

Replacing placeholders with real numbers

Once Part D and Part E complete, the README placeholders are replaced as follows:

Placeholder	Replace from
Table 1 (LIBERO-Long SR)	`outputs/eval/libero_long/eval_results.json` (config 07) and `outputs/eval/libero_long/config_06` (oracle)
Table 2 (ablation)	`paper_figures/ablation_v2/ablation_stats.csv`
§6.1 universality KS test	`paper_figures/universality/raw_distances.json`
§6.2 regret table	`paper_figures/regret_scaling/regret_vs_H.csv`
§6.3 triangulation P/R/F1	`paper_figures/triangulation/concordance_pr.json`
§6.4 detector comparison	`paper_figures/diagnostics/trigger_comparison.csv`
§6.5 boundary ratio	`paper_figures/diagnostics/boundary_ratio.json`
Inference cost latency/Hz	`paper_figures/diagnostics/inference_cost.json`

Theory and Method

A long-horizon manipulation policy executing action chunks faces a fundamental challenge: the flow-matching action distribution shifts sharply at phase boundaries (grasp → transport → place), but the policy must commit to a chunk of H actions before it can detect the shift. We call these boundaries Predictability Cliffs.

PACE v2 addresses this with three components:

Hierarchical Phase Encoder — a two-level FSQ (macro K₁ = 20, micro K₂ = 30) that makes phase structure identifiable across seeds via InfoNCE.
Three Cliff Estimators and Concordance — I^(1) (Bhattacharyya β_t), I^(2) (action-ensemble variance), I^(3) (velocity-field curvature) are fused via rank-based concordance C_t. A cliff is detected when all three estimators agree.
PACE Closed-loop Replanning (PCAR) — adaptive replanning triggered by C_t, with a budget constraint to prevent over-triggering.

Boundary-aware flow loss reweighting (w(β) = 1 + λ·β) further improves action prediction quality near phase transitions.

For the full mathematical derivation and ablation design, see docs/ARCHITECTURE.md.

Repository Structure

configs/
  train/          Three-stage training curriculum (01–03)
  ablation/v2/    Seven ablation configs + 2 sanity checks
baselines/        Cross-policy adapters (OpenVLA, π0, BC-ACT, Diffusion Policy)
lerobot_policy_phaseqflow/
  src/.../        Policy implementation (PhaseQFlowPolicy, PhaseQFlowConfig)
  phase_centric/  Training-side: phase posterior, cliff estimators (loss path),
                  PCAR / B-PCAR / boundary-aware flow loss
  inference/      Runtime-side (PACE v2): compute_policy_variance (I^(2)),
                  compute_velocity_curvature (I^(3)), ConcordanceDetector (C_t)
scripts/
  phenomenon/     Universality, regret scaling, triangulation experiments
  eval/           LIBERO-Perturbed and SimplerEnv evaluation
  figures/        Publication-grade figure generation
  diagnostics/    Boundary error, replan alignment, trigger comparison, cost
  calibration/    Concordance threshold and B-PCAR sweeps
tests/            Unit + smoke tests (covers training and inference paths)
docs/
  ARCHITECTURE.md      Full architecture specification
  OPERATIONS_GUIDE.md  Engineering handbook
  EXPERIMENT_RUNBOOK.md End-to-end GPU reproduction guide

The training and inference cliff signals are intentionally split:

phase_centric/cliff_estimators.py provides tensor-level functions (compute_I_hat_1/2/3, compute_concordance_C) consumed inside PhaseQFlowPolicy.forward() for loss / diagnostics.
inference/ exposes the runtime API used by evaluation rollouts (scripts/eval/libero_perturbed.py) — ConcordanceDetector + standalone estimator functions that read the policy's cached _last_beta and flow_action_head._last_condition to fuse I^(1)/I^(2)/I^(3) into C_t.

Citation

@inproceedings{pace2027,
  title     = {{PACE}: Predictability-Aware Closed-loop Execution
               for Long-Horizon Manipulation},
  author    = {[Authors]},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2027},
}

Name		Name	Last commit message	Last commit date
Latest commit History 71 Commits
baselines		baselines
configs		configs
docs		docs
lerobot_policy_phaseqflow		lerobot_policy_phaseqflow
scripts		scripts
tests		tests
.gitignore		.gitignore
LICENSE		LICENSE
MIGRATION_NOTES.md		MIGRATION_NOTES.md
README.md		README.md
REPRODUCIBILITY.md		REPRODUCIBILITY.md
environment.yml		environment.yml
pyproject.toml		pyproject.toml
requirements.txt		requirements.txt

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

PACE: Predictability-Aware Closed-loop Execution for Long-Horizon Manipulation

Implementation Status

Quick Start

Main Results

Table 1 — LIBERO-Long success rate: published baselines

Ablation Study (Table 2)

Phenomenon Results (Section 6)

§6.1 Universality — cliff occurrence across 4 policies (placeholder)

§6.2 Regret Scaling — δSR vs. chunk length H (placeholder)

§6.3 Triangulation Concordance (placeholder)

§6.4 Alternative Cliff Detectors (placeholder)

§6.5 Boundary Loss Ratio (placeholder)

Inference Cost (placeholder)

Verification & Reproduction Guide

Part A — CPU verification (no GPU, no data, ~5 min)

Part B — Prerequisites for real experiments

Part C — Three-stage training (real)

Part D — Evaluation pipeline (real)

Part E — Ablation, phenomenon, and figures (real)

Replacing placeholders with real numbers

Theory and Method

Repository Structure

Citation

About

Releases

Packages

Contributors

Languages

Folders and files

Latest commit

History

Repository files navigation

PACE: Predictability-Aware Closed-loop Execution for Long-Horizon Manipulation

Implementation Status

Quick Start

Main Results

Table 1 — LIBERO-Long success rate: published baselines

Ablation Study (Table 2)

Phenomenon Results (Section 6)

§6.1 Universality — cliff occurrence across 4 policies (placeholder)

§6.2 Regret Scaling — δSR vs. chunk length H (placeholder)

§6.3 Triangulation Concordance (placeholder)

§6.4 Alternative Cliff Detectors (placeholder)

§6.5 Boundary Loss Ratio (placeholder)

Inference Cost (placeholder)

Verification & Reproduction Guide

Part A — CPU verification (no GPU, no data, ~5 min)

Part B — Prerequisites for real experiments

Part C — Three-stage training (real)

Part D — Evaluation pipeline (real)

Part E — Ablation, phenomenon, and figures (real)

Replacing placeholders with real numbers

Theory and Method

Repository Structure

Citation

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages