Replan for Error Recovery: Training Embodied VLMs through Verifiable Interaction with Action-Grouped Turn-GSPO
Authors: Zhuoran Wang, Keqing Fan, Yang Li, Jinbo Su, Ke Wang, Sijia Luo, Yuanzheng Tan, Junkun Peng, Jing Zhang
Organizations: School of Information, Renmin University of China · Key Laboratory of Data Engineering and Knowledge Engineering, Beijing, China · Beihang University · Zphyture · Engineering Research Center of Database and Business Intelligence, Beijing, China
Embodied VLMs can execute increasingly complex manipulation tasks, yet execution errors can leave the environment in states that require replanning for recovery. Existing recovery methods either rely on inference-time mechanisms with fixed policies or learn from failure examples and reference corrections rather than verified action outcomes. To enable recovery learning through verifiable interaction, we introduce the Embodied Replanning for Recovery Environment (ERR-Env), which provides feedback on task progress and execution errors. We further construct ERR-Bench, a benchmark with held-out tasks spanning nominal, intermediate, and error states for evaluating recovery capability and execution reliability. We then propose Action-Grouped Turn-GSPO (AGT-GSPO), an online reinforcement learning method combining turn-level rewards for task progress and error states with action-grouped advantage estimation, and use it to train ERR-VLM, a compact 2B embodied VLM. Evaluation on ERR-Bench reveals a mismatch between task progress and execution reliability across diverse VLMs. AGT-GSPO improves ERR-VLM by 10.25 percentage points in Strict Success Rate for error-free completion and 22.95 percentage points in Recovery Rate, while ablations and zero-shot transfer further validate the effectiveness of our method.
Figures & tables
Figure 1: Recovery scenarios in ERR-Env. (a) A trajectory with errors and corrective replanning. (b) Examples of Manipulation Recovery, Occlusion Recovery, and Missing-Target Recovery.
Figure 2: Training process of AGT-GSPO. The policy interacts with ERR-Env over multiple turns and receives turn-level feedback on task progress and error states. Turns are grouped by action type for relative advantage estimation, enabling credit assignment among comparable recovery decisions.
Model
All Tasks
Recovery Rate
Recovery SSR
SR
PG
NPR
SSR
MR
OR
MTR
Total
MR
OR
MTR
Total
Closed-source models
GPT-5.6 Sol
40.98
76.81
51.97
15.44
38.92
53.29
55.09
49.10
14.97
11.98
8.38
11.78
Qwen3.8-Max
55.16
82.66
53.36
27.21
30.54
85.63
56.89
57.69
25.75
43.71
10.18
26.55
Qwen3.7-Plus
18.16
52.58
54.89
5.39
23.35
37.13
51.50
37.33
9.58
7.78
3.59
6.98
Open-source models
Table 1: Main results on ERR-Bench. We evaluate all tasks with four metrics: Success Rate (SR), Progress (PG), No-Penalty Rate (NPR), and Strict Success Rate (SSR). Recovery Rate and Recovery SSR are reported for Manipulation Recovery (MR), Occlusion Recovery (OR), Missing-Target Recovery (MTR), and their average (Total). All entries are percentages.
Method
All Tasks
Recovery Rate
Recovery SSR
SR
PG
NPR
SSR
MR
OR
MTR
Total
MR
OR
MTR
Total
SFT and full RL
ERR-VLM (SFT)
18.96
62.36
53.09
10.64
26.95
40.72
59.88
42.52
13.17
4.19
8.38
8.58
ERR-VLM (Full)
34.14
71.08
62.41
20.89
40.12
89.22
67.07
65.47
26.35
24.55
15.57
22.16
Recovery-data Ablations
w/o Recovery Data in RL
25.03
69.14
52.77
14.17
28.14
28.74
60.48
39.12
13.17
7.78
13.77
11.57
Table 2: Ablation results on ERR-Bench. Metrics follow Table 1 . All entries are percentages.
Model
RMBench
RoboVoLo
Observe and Pick Up
Cover Blocks
Place Block Mat
Memory
Common Sense
World Knowledge
Complex References
RoboBrain2.5-8B
65.0
33.3
38.3
62.5
95.5
20.0
45.0
ERR-VLM (SFT)
50.0
55.0
46.7
37.0
52.1
15.0
38.3
ERR-VLM (Full)
90.0
56.7
55.0
72.8
96.3
25.0
59.2
Table 3: Results on RMBench and RoboVoLo. Each task reports Progress in percentages.
Model
2 Targets
3 Targets
4 Targets
MR
MTR
Overall
PG
NPR
PG
NPR
PG
NPR
PG
NPR
PG
NPR
ARR
RoboBrain2.5-8B
47.5
20.0
48.3
70.0
25.7
20.0
55.0
10.0
72.5
40.0
8.9
ERR-VLM (SFT)
42.5
100.0
55.0
100.0
21.4
70.0
45.0
40.0
45.0
80.0
0.0
ERR-VLM (Full)
70.0
100.0
73.3
100.0
40.0
80.0
55.0
60.0
70.0
90.0
78.7
Table 6
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Setting
Base model
Qwen3-VL-2B-Thinking
Epochs
8
Learning rate
5×10−6
Batch size
64
Learning-rate schedule
Cosine
Warmup ratio
3%
Appendix
Table 5: SFT training configuration.
Parameter
Setting
Training batch size
8 prompts
Rollouts per prompt
6
Trajectories per update
48
Loss aggregation
Sequence-mean token-mean
Learning rate
2×10−6
Learning-rate schedule
Cosine decay
Appendix
Table 6: Online RL training configuration.
Model
Short-horizon
Long-horizon
SR
PG
NPR
SSR
SR
PG
NPR
SSR
Closed-source models
GPT-5.6 Sol
40.35
73.11
55.66
12.52
41.62
80.50
48.27
18.35
Qwen3.8-Max
65.38
83.25
59.25
30.23
44.95
82.06
47.47
24.20
Qwen3.7-Plus
21.30
52.57
57.26
3.33
15.03
52.58
52.53
7.45
Open-source models
Appendix
Table 7: Horizon-specific results on ERR-Bench. Short- and long-horizon episodes are reported separately. SR: Success Rate; PG: Progress; NPR: No-Penalty Rate; SSR: Strict Success Rate. All entries are percentages.
Method
Short-horizon
Long-horizon
SR
PG
NPR
SSR
SR
PG
NPR
SSR
SFT and full RL
ERR-VLM (SFT)
24.10
63.60
55.53
11.72
13.83
61.12
50.66
9.57
ERR-VLM (Full)
38.88
71.45
63.65
23.30
29.39
70.70
61.17
18.48
Recovery-data Ablations
w/o Recovery Data in RL
35.02
69.20
55.53
16.64
15.03
69.07
50.00
11.70
Appendix
Table 8: Horizon-specific ablation results on ERR-Bench. Short- and long-horizon episodes are reported separately. SR: Success Rate; PG: Progress; NPR: No-Penalty Rate; SSR: Strict Success Rate. All entries are percentages.
Model
Missing Tool Call
Parse Failure
Invalid Tool/Arguments
Total Error Turns
GPT-5.6 Sol
5
0
0
5
Qwen3.8-Max
744
51
29
824
Qwen3.7-Plus
0
0
0
0
Qwen3-VL-235B-A22B
0
0
3
3
InternVL3.5-8B
2434
1754
271
4459
Qwen3-VL-2B (Base)
3907
314
36
4257
Appendix
Table 9: Tool-call format errors observed during the main-result evaluation. Counts are error turns rather than unique trajectories.
Model
Error trajectories
State-changing executed trajectories
Later clean- format turn
All remaining turns malformed
Execution- affected rate
GPT-5.6 Sol
5
0
0
5
0.3%
Qwen3.8-Max
499
477
489
10
0.7%
Qwen3.7-Plus
0
0
0
0
0.0%
Qwen3-VL-235B-A22B
2
2
2
0
0.0%
InternVL3.5-8B
1304
1127
1244
60
4.0%
Qwen3-VL-2B (Base)
1351
404
1351
0
0.0%
Appendix
Table 10: Trajectory-level consequences of tool-call format errors in the main-result evaluation. Execution-Affected Rate is All Remaining Turns Malformed divided by 1,503 episodes.
Method
All Tasks
Recovery Rate
Recovery SSR
SR
PG
NPR
SSR
MR
OR
MTR
Total
MR
OR
MTR
Total
SFT and full RL
ERR-VLM (SFT)
18.96
62.36
53.09
10.64
26.95
40.72
59.88
42.52
13.17
4.19
8.38
8.58
ERR-VLM (Full)
34.14
71.08
62.41
20.89
40.12
89.22
67.07
65.47
26.35
24.55
15.57
22.16
Algorithm Ablations
Turn-GRPO
31.81
74.89
53.89
8.85
45.51
83.83
48.50
59.28
13.77
10.18
9.58
11.18
Appendix
Table 11: More ablation results on ERR-Bench. Metrics follow Table 1 . All entries are percentages.
Figure 4: Successful real-robot trajectories. (a) Two-target nominal manipulation. (b) Manipulation Recovery corrects a misplaced fruit before continuing the task. (c) Missing-Target Recovery completes feasible fruit placement while ignoring the unavailable bottle subtask.
Figure 5: Failure cases in real-robot execution. (d) Semantic or target errors lead to incorrect placements and task failure. (e) An occluded target leads to repeated same-point pick attempts and eventual failure.