Embodied-Corrector | Recover, Don’t Just Continue

RESEARCH FILM

Recover, don’t just continue.

Why physical failures need causal diagnosis, how we train RynnBrain, and what changes during recovery.

Method diagrams and continuous simulator excerpts. Demonstrations use a GPT proxy with state-assisted control; quantitative results are reported separately from the manuscript.

01 / THE RESEARCH IN BRIEF

Abstract

Reliable embodied execution requires more than generating a plausible plan: an embodied brain must understand physical causes of failure, adapt its guidance to the executor’s capabilities, and use interaction history to revise subsequent decisions. We introduce Embodied-Corrector, a corrective brain trained from RynnBrain-8B to connect observed outcomes with causal diagnosis and executable recovery.

We construct controlled failure–recovery branches from successful trajectories, changing one physical factor at a time. Evaluation-first SFT first teaches outcome judgment and cause analysis, then recovery planning. Noise-Aware Entropy-Gated Policy Optimization (NEPO) combines group-relative rewards with token-level guidance across generated and verified evaluation contexts. At deployment, the model produces Judge → Cause → Sub → Fine from public execution evidence and within-episode history.

The manuscript evaluates failure reasoning, planning transfer, and recovery with fixed execution backends. It reports 55.0% → 68.9% macro-average success with a fixed code backend and 1.2% → 7.3% success for a frozen π_0.5 on RoboCasa365 Composite-Unseen. These simulation results connect improved diagnosis with more effective recovery while keeping the executor fixed.

02 / WHY RECOVERY MATTERS

Motivation

A capable brain must know what went wrong.

Physical execution is constrained by the embodiment, controller, and environment. How can an embodied brain explain unexpected outcomes and use feedback to plan a recovery its current executor can realize?

01

Understand the cause

Locate the action, relation, or missing prerequisite that first invalidated progress—not just the unsuccessful outcome.

Physical causal understanding

02

Adapt to the executor

Translate a recovery into state-conditioned subtasks that the current controller or frozen VLA can actually realize.

Capability-aware planning

03

Learn from the interaction

Use prior actions, observations, and corrections as evidence for the next decision, without updating model parameters.

In-context adaptation

“It failed” is not enough. On EB-ALFRED, cause-level feedback improves replanning more consistently than repeating the failed outcome.

**Figure 1. Failure diagnosis and feedback controls.** Frozen multimodal models struggle to identify physical causes on EB-ALFRED. Cause-level feedback improves replanning more than repeating the unsuccessful outcome.

03 / HOW WE TRAIN THE CORRECTIVE BRAIN

Method

Assess first. Then plan a recovery.

We train RynnBrain-8B with evaluation-first SFT and NEPO. A single inference connects execution evidence to four structured outputs, giving the same downstream planner a better basis for its next action.

RynnBrain-8BEvaluation-first SFTNEPOEmbodied-Corrector

01

Judge

Did the subtask succeed?

Placement incomplete

02

Cause

Which condition was violated?

The carton was never grasped

03

Sub

What should happen next?

Locate and pick up the carton

04

Fine

How can it be executed?

Grasp → verify → lift → place

Illustrative summary of the paper’s orange-juice recovery case.

DATA CONSTRUCTION

One controlled change. A learnable physical cause.

Branch successful trajectories at subtask boundaries, intervene on one factor, and pair failures with aligned correction targets. Keep only failures that are observable and recoverable.

NavigationReachabilityGrasp poseTimingArticulationPrerequisitesTarget grounding

Sources: RoboCasa365 and OmniBehavior. Source trajectories are split before perturbation generation to prevent shared prefixes across training and evaluation.

Paper diagram of successful trajectory sources, controlled perturbations, failed and corrected traces, and four-field training targets. — FIGURE 2 Controlled failure–recovery data

EVALUATION-FIRST TRAINING

Turn verified evaluations into the model’s own reasoning.

First learn Judge + Cause, then learn Sub + Fine conditioned on verified evaluations. NEPO combines group-relative rewards with entropy-gated on-policy distillation to reduce the training-to-deployment mismatch.

1Learn to assess before learning to continue.
2Retain positive guidance; attenuate uncertain negative signals.
3Generate all four fields in one deployment pass.

Verified evaluation labels are available during training only.

Paper diagram of evaluation-first supervised training, student and verified-evaluation teacher views, entropy gating, and the GRPO branch of NEPO. — FIGURE 3 SFT → NEPO → one-pass correction

INSIDE NEPO

Same policy. Same planning tokens. Two evaluation contexts.

Training can use a verified evaluation, while deployment must rely on the model’s own assessment. NEPO turns this context difference into a token-level learning signal.

Δ = q − p

p · probability under the generated evaluation q · probability under the verified evaluation

Keep positive guidanceGate uncertain negative guidanceCombine with GRPO rewards

Verified evaluations are training-only. Deployment generates all four fields.

THE CLOSED LOOP

Better feedback. The same execution backend.

The corrective brain reads public execution evidence and returns a structured correction. A fixed planner revises the workflow; the environment verifies the result.

Read the Harness details

Closed-loop Harness connecting task and context, fixed execution interfaces, Embodied-Corrector, and workflow revision. — **Figure 6. Corrective Harness.** A fixed executor receives structured feedback. Formal evaluation uses within-episode working memory only.

04 / FROM DIAGNOSIS TO EXECUTION

Experiments

Better correction, with fixed executors.

Evaluated at the brain level and with fixed execution backends. Explore the reported results across four complementary settings.

FIXED CODE BACKEND

55.0 → 68.9%

+13.9 percentage points

Unweighted mean across 4 OmniBehavior and 6 LIBERO-PRO conditions.

Table 3 · Explore results ↗

FROZEN π_0.5

1.2 → 7.3%

+6.1 percentage points

RoboCasa365 Composite-Unseen, with the low-level policy frozen.

Figure 4 · Explore results ↗

PLANNING TRANSFER

30 → 78%

+48 percentage points

EB-ALFRED Base, without an explicit failure prompt or corrective Harness.

Table 2 · Explore results ↗

EmbodiedDiag-Bench

Execution understanding and corrective planning across navigation and manipulation.

Dataset & evaluation code ↗

Navigation

Accuracy (%) ↑

Model	Judge	Cause	Sub	Fine
GPT-5.5	52.8	39.8	72.7	39.4
RoboBrain2.5-4B	31.5	12.2	26.1	21.2
RynnBrain-8B	23.6	13.6	23.4	19.4
InternVL3-8B	48.3	26.2	21.8	17.7
MiMo-Embodied-7B	30.1	16.9	31.6	25.3
VL-7B baseline	52.5	23.8	23.2	19.3
VL-32B baseline	49.5	29.2	48.7	39.8
RynnBrain-SFT	75.9	53.4	68.6	66.7
RynnBrain-OPD	79.0	57.0	70.0	68.0
RynnBrain-GRPO	74.2	51.5	73.0	69.2
RynnBrain-NEPO (ours)	79.6	58.2	72.6	71.6
w/o Entropy Gate	73.1	50.8	69.9	68.5

Manipulation

Accuracy (%) ↑

Model	Judge	Cause	Sub	Fine
GPT-5.5	57.7	43.2	65.4	41.1
RoboBrain2.5-4B	33.5	12.3	29.7	24.1
RynnBrain-8B	31.9	18.7	31.6	25.8
InternVL3-8B	40.1	17.9	32.1	26.8
MiMo-Embodied-7B	22.2	19.3	35.6	27.6
VL-7B baseline	33.7	13.5	28.8	26.0
VL-32B baseline	54.6	33.2	42.9	34.3
RynnBrain-SFT	80.7	51.4	62.5	61.6
RynnBrain-OPD	80.0	54.0	65.4	61.0
RynnBrain-GRPO	76.3	49.6	63.6	62.1
RynnBrain-NEPO (ours)	81.2	54.6	64.9	63.7
w/o Entropy Gate	75.0	48.7	63.1	61.4

Manuscript Table 1. Judge: execution outcome · Cause: failure diagnosis · Sub: next subtask · Fine: grounded action details. All values are reported in the manuscript.

**Figure 4. Frozen-VLA success on RoboCasa365.** Corrective planning improves composite tasks, while Atomic-Seen success decreases. Values are percentages reported in the manuscript.

POST-TRAINING DYNAMICS

Stronger reward. Without entropy collapse.

Under matched VQA post-training conditions, NEPO achieves the highest answer-level reward while avoiding the low-entropy regime of vanilla OPD.

Original loss, answer reward, and token entropy curves comparing GRPO, GSPO, OPD, and NEPO. — **Figure 5. Post-training dynamics.** Loss, answer reward, and token entropy under matched VQA post-training conditions.

FAILURE → DIAGNOSIS → RECOVERY

12 distinct tasks, 12 successful recoveries

6 different LIBERO-PRO tasks and 6 different RoboCasa365 tasks. Each RoboCasa task combines navigation and manipulation. Watch the visible error, the English Judge → Cause → Sub → Fine explanation, and the successful correction.

LIBERO-PRO 6 tasks

Back to directory ↑

01

Put the black bowl in the bottom drawer of the cabinet and close it

high grasp + partial drawer closure with hand withdrawal · 1×

Recovery succeeded

02

Put the yellow and white mug in the microwave and close it

lateral grasp offset + incomplete door closure · 1×

Recovery succeeded

03

Turn on the stove and put the moka pot on it

missed switch contact + placement offset · 1×

Recovery succeeded

04

Put both the cream cheese box and the butter in the basket

lateral grasp offset + placement outside basket · 1×

Recovery succeeded

05

Put the white mug on the left plate and put the yellow and white mug on the right plate

high grasp + placement on wrong plate · 1×

Recovery succeeded

06

Pick up the book and place it in the back compartment of the caddy

lateral grasp offset + placement outside organizer · 1×

Recovery succeeded

RoboCasa365 6 tasks

Back to directory ↑

07

Place the fruits in the fridge on the cutting board by the sink to make a fruit bowl. Leave the non-fruit items in the fridge.

wrong-direction navigation + missed fruit grasp · 1×

Recovery succeeded

08

Transport the chicken drumstick from the fridge to the microwave and turn on the microwave.

high grasp + incomplete microwave door closure · 1×

Recovery succeeded

09

Pick the mug from the cabinet, place it under the coffee machine dispenser, and press the start button.

wrong-direction navigation + missed start button · 1×

Recovery succeeded

10

Pick up the cup and bowl from the counter, place them in the dishwasher, and close the dishwasher door.

high grasp + premature door-closing withdrawal · 1×

Recovery succeeded

11

Pick the tangerine, cucumber from the counter and place them in the cabinet. Then close the cabinet.

wrong-direction navigation + high grasp · 1×

Recovery succeeded

12

Move the pot and pan from the stove into the sink.

lateral handle-grasp offset + wrong-direction navigation · 1×

Recovery succeeded

05 / CITATION

Citation

If this work supports your research, please cite the arXiv paper.

@misc{pan2026recover,
  title         = {Recover, Don't Just Continue: Learning from Feedback for Physical Failure Understanding and Correction},
  author        = {Pan, Miao and Yao, Kaixiang and Zhou, Xiaohe and Wang, Xu and Li, Qinfeng and Zhang, Wenqi and Zhang, Xuhong},
  year          = {2026},
  eprint        = {2611.15651},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2611.15651}
}

This entry cites the arXiv paper at arXiv:2611.15651.