High fluctuation in training loss during SFT on RoboTwin tasks - Is this expected?

[lwbscu]

Hi LingBot-VLA team,

Thanks for releasing this amazing work and the highly efficient training codebase!

I am currently running the Post-Training (SFT) pipeline using the provided lingbot-vla-4b model and Qwen2.5-VL-3B-Instruct base. I have conducted rigorous experiments on both a single-task dataset (e.g., click_bell) and the RoboTwin 5-tasks mixed dataset.

1. My Training Setup

• Hardware: 8x A800 GPUs
• Batch Size: micro_batch_size = 4 per GPU (resulting in global_batch_size = 32)
• Config: Modified from robotwin_load20000h.yaml

2. Observation A: High Loss Fluctuation

During training, I plotted the loss from loss.jsonl. Interestingly, even in the single-task training scenario (as shown in the attached figure for the click_bell task), the raw training loss (blue line) fluctuates heavily with dense spikes.

3. Observation B: Significant Evaluation Performance Gap

While the official claims demonstrate a highly capable model (e.g., ~80%+ success rates on RoboTwin tasks), my evaluation results fall far short of this, despite strictly following the SFT pipeline:

• Single-Task SFT: When training exclusively on the simplest task (click_bell), the success rate is only around 30%.
• Multi-Task SFT (5 Tasks): I then strictly followed the official documentation snippet: "Training Configuration: We provide the mixed post-training configuration in five RoboTwin 2.0 tasks ('open microwave' 'click bell' 'stack blocks three' 'place shoe' 'put object cabinet')." Surprisingly, when training on the 5-task mixed dataset, the success rate for individual tasks drops drastically to single digits (just a few percent).

4. My Questions:

1. Loss Fluctuation: Given the 8-GPU setup (micro_batch_size=4), is this high-frequency fluctuation completely normal for LingBot-VLA's architecture during SFT, even on a single task? How do you usually determine optimal convergence?
2. Performance Reproduction: How can we reproduce the high success rates mentioned in your reports? Are there specific hyperparameters, data augmentation strategies, or longer training epochs required that are not reflected in the default robotwin_load20000h.yaml?
3. Multi-task Degradation: Is it expected that the 5-task mixed training performs significantly worse than single-task training under the current config? Does the multi-task setup require a different sampling strategy or a much larger batch size to handle task interference?

Thanks in advance for your time and insights!

[YangS03]

The performance is not normal. Can you provide a complete rollout video?

[LuFan31]

Thank you for your interest in LingBot-VLA!

1. We utilize L1 flow-matching loss for training on RoboTwin, which inherently produces larger numerical values and more significant fluctuations compared to MSE flow-matching loss. However, L1 fm loss achieves better performance with fewer training steps in RoboTwin.
   Regarding convergence, a global batch size of 32 combined with only 8k training steps is, in our experience, far from sufficient to reach optimal convergence on RoboTwin. We recommend utilizing multi-node training to increase the global batch size and significantly extending the total number of training steps.

2. The robotwin_load20000h.yaml file is primarily intended for ablation studies during our pretraining. The configuration used for the high success rates mentioned in our reports is not identical to this default setting.
   To assist the community in reproducing our results, we plan to open-source the weights finetuned on 50 RoboTwin tasks, along with the corresponding evaluation code soon. Please stay tuned for these updates.

3. We have not conducted experiments specifically comparing single-task versus multi-task performance degradation. In our RoboTwin experiments, both Pi0 and LingBot-VLA adopt a multi-task mixed training setting. Our primary goal is to evaluate the model's generalizability across diverse scenarios rather than its ability to overfit a single specific task.

[lwbscu]

ab7e46983f8189a5c9abdcb3ccfba695.mp4

b58174efd94da2c13fdbe3d6c29ef7a6.mp4

295a31d3f48ba4b6505130a0b736b816.mp4

28f930b9756ea71efe78acde55ce1571.mp4

c66d8370cdd81b4ecb60932796596560.mp4

性能不正常。你能提供一个完整的推广视频吗？

[gzyabc]

ab7e46983f8189a5c9abdcb3ccfba695.mp4
b58174efd94da2c13fdbe3d6c29ef7a6.mp4
295a31d3f48ba4b6505130a0b736b816.mp4
28f930b9756ea71efe78acde55ce1571.mp4
c66d8370cdd81b4ecb60932796596560.mp4

性能不正常。你能提供一个完整的推广视频吗？
Hi, I am encountering the same issue. I'm using 800 data samples collected from the Agibot G2 robot and following the official training configuration.

However, the memory consumption during training is massive—it requires two A100 GPUs (160GB total). I've also noticed that the training loss fluctuates significantly. Even after training for 150,000 steps (which takes about 3 to 4 days), the actual deployment performance is very poor. It feels like the model hasn't truly 'learned' the tasks.Would you be open to sharing your contact information so we could discuss this further

[zxduan-r]

Thanks for the official clarification! We'd like to share our reproduction results as additional reference for the community.

Our Setup

We also trained on the 5-task mixed dataset using the robotwin_load20000h.yaml config, but with a larger global batch size via multi-node training:

Parameter	Value
Hardware	2 nodes × 8 GPUs (16 GPUs total)
micro_batch_size	32
global_batch_size	512
lr	1e-4
num_train_epochs	69
Effective steps	10,005
loss_type	L1_fm
Base model	lingbot-vla-4b

Training loss converged from ~0.53 → ~0.011.

Evaluation Results (unseen instructions, 100 trials, demo_clean)

Task	Success Rate
open_microwave	75/100 = 75.0%
click_bell	66/100 = 66.0%
place_shoe	48/100 = 48.0%
stack_blocks_three	26/100 = 26.0%
put_object_cabinet	22/100 = 22.0%
Average	47.4%

Compared to the results reported in issue #21 (single-digit % with GBS=32), increasing the global batch size to 512 with multi-node training seems to significantly improve performance. This is consistent with the team's suggestion to increase batch size and training steps.

That said, there is still a noticeable gap from the paper's reported ~80%+. We look forward to the release of the fine-tuned weights and the exact training configuration used in the paper.

Also tagging the related discussion in #24 for cross-reference.

[gzyabc]

@zxduan-r However, we don't have access to such high-end computing resources. Given that we only have two A100 GPUs, is it still possible to achieve this level of performance?

[gzyabc]

@zxduan-r However, we don't have access to such high-end computing resources. Given that we only have two A100 GPUs, is it still possible to achieve this level of performance?

I have increased the batch size, and the loss is now decreasing normally. I plan to complete the training with this configuration first and then evaluate the actual performance on the robot.

[MsterDC]

Thank you for your interest in LingBot-VLA!

1. We utilize L1 flow-matching loss for training on RoboTwin, which inherently produces larger numerical values and more significant fluctuations compared to MSE flow-matching loss. However, L1 fm loss achieves better performance with fewer training steps in RoboTwin.
   Regarding convergence, a global batch size of 32 combined with only 8k training steps is, in our experience, far from sufficient to reach optimal convergence on RoboTwin. We recommend utilizing multi-node training to increase the global batch size and significantly extending the total number of training steps.
2. The robotwin_load20000h.yaml file is primarily intended for ablation studies during our pretraining. The configuration used for the high success rates mentioned in our reports is not identical to this default setting.
   To assist the community in reproducing our results, we plan to open-source the weights finetuned on 50 RoboTwin tasks, along with the corresponding evaluation code soon. Please stay tuned for these updates.
3. We have not conducted experiments specifically comparing single-task versus multi-task performance degradation. In our RoboTwin experiments, both Pi0 and LingBot-VLA adopt a multi-task mixed training setting. Our primary goal is to evaluate the model's generalizability across diverse scenarios rather than its ability to overfit a single specific task.

When will the post-training configuration be released for all tasks on robotwin 2.0? I believe this is crucial for reproducing the full results of lingbot-vla.