Summary
This week's theme connects deployment-time behavior adaptation, selective experience memory, and video-based planning as complementary approaches to robust robot control. RynnUnify adds current-week evidence for integrating embodied reasoning, future-state prediction, and action generation in multi-step tasks, while progress papers extend transfer, force-reactive control, and feedback-driven policy revision.
Situation
The representative Introductions identify complementary barriers to robot deployment: unexpected conditions that require adaptation without human resets, partial observability that requires long-term recall, and mismatches between broad pretrained knowledge and environment-specific dynamics. Adapt On-the-Go addresses the first barrier by selecting and optionally fine-tuning pretrained behaviors within a single deployment episode. MemER uses a hierarchy that retains task-relevant visual keyframes and issues subtasks to a low-level controller, avoiding the cost and generalization risks of conditioning on full observation histories. Solving New Tasks by Adapting Internet Video Knowledge studies how in-domain information can adapt internet-trained video models for text-conditioned planning and policy supervision.
Current-week evidence from RynnUnify: Towards Unified Embodied Brain (2610.32965v1) extends this picture toward integrated systems. It interleaves language reasoning with video prediction and connects these capabilities to control through latent planning and joint video-action training. Its Introduction reports simulated and real-robot multi-step manipulation, while identifying accumulated prediction errors and long-horizon consistency as continuing challenges. Together, this evidence points to the need to coordinate reusable behaviors, selective memory, and environment-grounded predictions rather than rely on pretrained competence alone.
Infographic (English)

Progress
Cross-Controller Transfer via In-Context Behavior Adaptation <See Details on Fugu-MT>
This work studies transferring robot data across embodiments and controllers without letting differing action interpretations interfere with learning. It extends behavior adaptation to controller-dependent action semantics.
Self-Labeled Preference Transfer for Steerable Robot Manipulation <See Details on Fugu-MT>
Self-Labeled Preference Transfer adapts learned manipulation preferences to new tasks and objects while maintaining task success. It targets the gap where preference-trained policies can follow new-task preferences but achieve limited task success.
Learning Force-Reactive Humanoid Loco-Manipulation <See Details on Fugu-MT>
FLORE combines a force-reactive diffusion policy with a hybrid-compliance whole-body controller for humanoid loco-manipulation. It extends environment-grounded control to measured interaction forces while maintaining balance.
When Human Corrections Travel: Safety in Federated VLA Adaptation <See Details on Fugu-MT>
This work examines how adaptations derived from human corrections can be shared safely across distributed robots through federated VLA learning. It discusses requirements for preserving correction context and uncertainty, closed-loop evaluation of candidate updates, and leakage assessment, beyond adaptation on a single robot.
RoboDoctor: Training-Free Recursive Self-Improvement for Robot Policies <See Details on Fugu-MT>
RoboDoctor links diagnosis, grounded actions, and persistent experience in a training-free framework for improving robot policies. It uses execution feedback to revise later decisions while keeping model weights fixed.
Outlook
Outlook Summary
Deployment-time adaptation may become cumulative, expanding pretrained behaviors through persistent execution feedback and shared human corrections. Cross-controller transfer and preference adaptation make correct action interpretation and preserved task success important tests, but reliable lifelong adaptation remains unproven. Longer tasks also require selective memory and physically grounded predictions. MemER proposes deleting keyframes and reducing inference latency, alongside broader sensing and mobile multi-room tests. Adapt2Act points toward cheaper, suboptimal training data and repeated observations collected under the current policy, while calling for explanations of differing planning and policy-training results. Longer executions should test whether these methods sustain control rather than merely improve isolated tasks.
Infographic (English)

Three-Year Movement
Over roughly three years, this scenario would organize cumulative robot adaptation around a tested operating range rather than assume that combining promising methods produces reliable long-duration control. Existing work supplies separate foundations: adaptation of available behaviors, memory across minutes, and better sensing during interaction. This week's work connects those foundations to predicted futures and persistent corrections, but does not establish reliable combined performance.
The near-term setup is to define which conditions a particular robot, task, and controller can handle. The mechanism comes from aviation flight testing: probe the boundaries of that operating range and expand it only when repeated tests support the change. This gives deployment experience a disciplined role in deciding where an adapted system can be used.
By around year three, the central question would be whether accumulated evidence supports a broader tested range during longer executions. The monitoring cue is repeatable performance under the conditions proposed for inclusion, rather than success in an isolated demonstration. The main caveat remains the evidence gap: the supplied work suggests an integration direction, not proof that its components produce dependable cumulative adaptation.
Over roughly three years, this scenario would develop an execution journal that separates intended effects from observed outcomes, so a prediction alone cannot justify a later robot action. Borrowing from transactional database design, the journal would track which observations later subtasks depend on and require enough execution evidence before treating a proposed effect as established. In year one, it would initially observe an existing policy without changing control, then be tested on failed grasps and observations that become outdated or matter much later. Once failures are understood, the wrapper could request an inspection, delay a dependent action, or ask for help while leaving existing safety controls intact.
Progress would require better task success or fewer interventions across at least two task families with delayed dependencies, compared with retrieval and feedback baselines under matched resource limits. Those comparisons must count inspection time and engineering effort, keep storage bounded, and avoid more missed control deadlines rather than hide the cost of added checks. If that threshold is met, year two would turn journal entries into reusable corrections that specify when they apply and what evidence should confirm their expected effect. Corrections would first be reviewed locally, and another robot would have to verify the physical result in its own setting before relying on them.
By year three, successful systems could attempt longer supervised manipulation sequences and limited movement between work areas, where outdated observations and delayed evidence become harder to manage. Memory would need to retain rare but important observations without growing indefinitely, while new behavior variants would require physical outcome checks and tests that earlier capabilities still work. A mismatch would reveal a missing condition for reuse, allowing corrections to become more selective and potentially reducing repeated errors and operator interventions over subsequent runs.
Monitoring would therefore track intervention rates alongside false confirmations and failures to recognize that earlier evidence no longer applies, while obsolete corrections are retired. This path would weaken if gains disappeared after accounting for extra sensing, small physical changes broke corrections, or ordinary retrieval achieved comparable results more cheaply. The journal could block an unsupported next action, but it could not undo physical damage or establish general control competence beyond tested conditions.
Over roughly three years, this scenario would focus cumulative adaptation on a bounded, recurring task: restarting a manipulation cell after maintenance changes its operating conditions. A gripper replacement, fixture movement, or controller change gives technicians a concrete reason to check both physical operation and their account of what changed. This provides a setting where deployment experience could become useful without assuming a general lifelong learning system.
The near-term foundation remains a set of separate capabilities, including refinement of existing behaviors and retrieval of observations across minutes. Contact sensing adds physical evidence, while this week's work points toward controller-aware transfer and persistent execution feedback. The mechanism is to bring these capabilities together around maintenance restarts, where corrections must relate to the changed setup rather than merely repeat an earlier successful action.
At around year three, the scenario's relevant outcome would remain accurate, dependable restart support within this bounded setting, rather than unrestricted adaptation. A practical monitoring cue is whether restored operation agrees with the technician's account of the changes. The central caveat is that the supplied evidence does not yet demonstrate a reliable combined system or establish a year-by-year path to that outcome.
1-Year / 3-Year Research-Application Infographic

References
- Adapt On-the-Go: Behavior Modulation for Single-Life Robot Deployment - Authors: Annie S. Chen, Govind Chada, Laura Smith, Archit Sharma, Zipeng Fu, Sergey Levine, Chelsea Finn / <See Details on Fugu-MT> / License: CC-BY-4.0
- Solving New Tasks by Adapting Internet Video Knowledge - Authors: Calvin Luo, Zilai Zeng, Yilun Du, Chen Sun / <See Details on Fugu-MT> / License: CC-BY-4.0
- MemER: Scaling Up Memory for Robot Control via Experience Retrieval - Authors: Ajay Sridhar, Jennifer Pan, Satvik Sharma, Chelsea Finn / <See Details on Fugu-MT> / License: CC-BY-4.0
- RynnUnify: Towards Unified Embodied Brain - Authors: Minghao Zhu, Bohan Hou, Kun Xiang, Zixuan Wang, Yinzhou Tang, Yifei Yang, Zhikai Wang, Tianyi Zhang, Kehan Li, Haoyu Zhao, Siteng Huang, Leigang Qu, Sicong Leng, Dongchi Huang, Mingxiu Chen, Shijian Lu, Yue Wang, Xiaodan Liang, Zhenyu Zeng, Deli Zhao, Chengju Liu, Jianfei Yang, Xin Li / <See Details on Fugu-MT> / License: CC BY 4.0