---
license: apache-2.0
tags:
  - robotics
  - world-model
  - robotwin
  - chronoloop
---

# ChronoLoop on 5 hard RoboTwin 2.0 tasks (epoch 5)

Epoch-5 checkpoints of two ChronoLoop runs (the `chrono` branch of
[anhdao69/LoopWAM_NT](https://github.com/anhdao69/LoopWAM_NT)) trained on five RoboTwin 2.0 tasks:

| Folder | Run | Memory |
|---|---|---|
| `CL-0_no-memory_a4/` | CL-0, the no-memory control | none (stateless) |
| `CL-A_mem16-learned-loopwrite_a4/` | CL-A | 16 memory tokens, learned source, loop write, TBPTT S = 4 |

Both use LoopWAM-S v0 (3 pre + 6 core x 4 loops + 3 coda), K_a = 4, initialized from the full-suite LIBERO v0 4/4
parent ([anhdao69/LoopWAM_NT](https://huggingface.co/anhdao69/LoopWAM_NT)) with new 14-D action and proprio interfaces.

**These are mid-schedule checkpoints.** The runs follow a 10-epoch cosine schedule (70,220 updates); epoch 5 ends at
update 35,111, where the learning rate is still 53% of its peak.

## Data

Tasks: hanging_mug, blocks_ranking_size, scan_object, place_can_basket, handover_mic — the five tasks with the lowest
trimmed-mean success in LiLa-WAM (arXiv 2608.03701) Table 8. 550 demonstrations per task (50 clean + 500
randomized) from the FastWAM RoboTwin release: 2,750 episodes, 792,168 training windows. FastWAM RoboTwin input
protocol: 9 frames (offsets 0..32 step 4) and 32 actions per window, three cameras packed into 384x320, 14-D joint
actions and state, global z-score with the published FastWAM statistics (`data/dataset_stats.json`).

## Files

- `<run>/epoch_05/policy.pt` — weights only (`chronoloop-v1`), FP32.
- `<run>/epoch_05/config.json` — memory flags, parent and schedule hashes, the epoch's last update.
- `<run>/manifest.json`, `metrics.jsonl`, `timing.json` — the full training record (per-update losses, LR, gates).
- `data/dataset_stats.json` (normalization), `data/data_manifest.json` (episodes, windows, protocol).

Load with `fastwam.models.wan22.chronoloop_robot.create_chronoloop_robot(cfg, vae_path, action_dim=14,
proprio_dim=14, checkpoint_path=".../policy.pt")`, where `cfg` is the `ChronoConfig` stored in the checkpoint.

## Training

LoopWAM-S v0, K_a = 4; AdamW (0.9, 0.95), weight decay 0.01; LR 3e-5 backbone, 3e-4 memory and robot interfaces;
3% warmup and cosine to 1% over 10 epochs (70,220 updates of 32 streams x 4 windows); TBPTT over 4 windows with
stride 24. Two H100s per run.

| Epoch | CL-0 loss (video / action) | CL-A loss (video / action) |
|---|---|---|
| 1 | 0.1014 / 0.1139 | 0.1014 / 0.1139 |
| 2 | 0.0756 / 0.0170 | 0.0756 / 0.0170 |
| 3 | 0.0691 / 0.0113 | 0.0690 / 0.0110 |
| 4 | 0.0651 / 0.0085 | 0.0648 / 0.0072 |
| 5 | 0.0623 / 0.0067 | 0.0618 / 0.0051 |

Per-epoch means of the training loss. CL-A's memory read gates start at zero, so the two runs coincide for two epochs;
CL-A's action loss is 24% lower in epoch 5.

## RoboTwin 2.0 evaluation

RoboTwin 2.0 harness (`script/eval_policy.py`, aloha-agilex), unseen instructions, 50 episodes per task in each
setting (scene seeds from 100000, kept only if the expert planner solves them; step limits from RoboTwin's task
configs). The policy uses the training input protocol, executes the first 24 of 32 predicted joint targets and replans;
10 Euler steps, no CFG. CL-A's memory resets each episode.

Mean success over the 5 tasks (mean of per-task rates):

| Run | Clean | Randomized |
|---|---:|---:|
| CL-0 (no memory) | 55.2% | 50.0% |
| CL-A (mem16) | **57.6%** | **53.1%** |

| Task | CL-0 Clean | CL-A Clean | CL-0 Randomized | CL-A Randomized |
|---|---:|---:|---:|---:|
| hanging_mug | 12/50 | 15/50 | 15/50 | 14/49 |
| blocks_ranking_size | 24/50 | 23/50 | 18/50 | 18/50 |
| scan_object | 30/50 | 29/50 | 22/50 | 25/50 |
| place_can_basket | 22/50 | 27/48 | 21/50 | 25/49 |
| handover_mic | 50/50 | 49/50 | 49/50 | 50/50 |

(CL-A's evaluation ran each job as two halves; in three jobs one or two seeds were run by both halves and count once.)

On the 220 scenes per setting that both models ran, CL-A leads by 1.3 points on Clean and 5.4 on Randomized; neither
difference is statistically significant (exact McNemar p = 0.80 and 0.14). Per-seed outcomes:
`eval/robotwin5hard_epoch05_results.json`. These models are trained on 5 tasks only, so their numbers are not
comparable to multi-task results on all 50 RoboTwin tasks.

Report: `plans/performance/ChronoLoop_Robotwin5hard_Epoch5_20261017.md` on the `chrono` branch (training, evaluation protocol, harness changes, environment and commands).