Long Paths and
Shortcuts to
Understanding VLA

Experiments in Planning, Reward Design,
and Adaptation to a New Robot

Alexei Kornaev, Dr.Sc., Prof.

Innopolis University

Physical AI & VLM
Lomonosov Academic Exchange Day
30 October 2026 · Moscow

Our long-term goal is useful robots for people and the environment. Today I will show two routes for adapting existing models: guidance above a fixed controller, and learning for a different robot.

Title illustration: supplied robot-at-a-crossroads artwork.

Our path to embodied robotics

Embodied
roboticsLong-term ambition

↓Set the robots free

↓Let them work hard

↓Let them hand over
their salaries to us

↓Let them take care of us
and the environment

Physical understanding
+ uncertaintyFuture research:
physics-informed methods
PINN/iPINN · JEPA
knowing when to ask

→

Test on
our tasksLimited perception
Complex goals
Changed embodiment

→

Reproduce existing
VLA systemsWithin available compute

→

Our current focus

High-level planningGuide a fixed VLA executor
without retraining it

Reinforcement learningAdapt the policy in simulation
to a new robot: CDPR

Initial: Let me begin at the end: embodied robotics. Clicks 1–4: We want to set the robots free, let them work hard, let them hand over their salaries to us [pause]—and let them take care of us and the environment. Click 5: That is our long-term ambition. Physical understanding and knowing when to ask for help are directions we want to explore. Click 6: Our current work is testing existing systems on our own tasks. Click 7: We start with what we can reproduce within our available compute. Clicks 8–9: Today I will focus on two directions: high-level planning above a fixed VLA executor, and reinforcement learning to adapt a policy in simulation to a new robot. Allow approximately 60–90 seconds; rehearse the reveal sequence. The ambition is distinct from demonstrated experimental results.

The people behind the experiments

Alexei KornaevScientific supervisor

Alexander MaloletovHead, NTI Center*

Damir NurtdinovVLA / RL, CDPR

Denis DegtyarevPlanner evaluation

Zlata MishustovaWheeled robot

Kamil NurievRewards, physical criteria

Egor NovikovVLA adaptation, rewards

Artem TuzovData and tooling

Yuri KazakovCartography, control

*NTI Center for Robotics and Mechatronics Component Technologies

Our team combines planner evaluation, VLA adaptation, control, and experimental tooling. The labels indicate research areas rather than authorship of every paper.

Part 1 · Planning above a fixed controller

1. 01Planning above a fixed controller
2. 02Learning for a new embodiment
3. 03What comes next

Can better guidance help a controller that we do not retrain?

Can better guidance help a controller that we do not retrain?

Planner pipeline: manipulation

Manipulation study: planner-selected subgoals for a fixed SmolVLA executor.

Our contribution: measuring help, harm, and chance

Count what a net success rate hides.
Replay each task from the same start, with and without guidance; count rescued failures and disrupted successes separately.

Measure how much change happens without a planner.
Replay the fixed policy against itself to estimate its own outcome variability.

Test the cost and limits of intervention.
Compare planner models, reference goals and larger execution budgets to examine why progress may improve while completion does not.

What can the planner change?

“Put the three food boxes into the organizer.”

PlannerInstruction + observation↓ current goal

“Pick up the butter.”

↓Fixed SmolVLAImages + state + goal → actions

Pick / place cream cheese → Pick / place popcorn → Pick / place butter

The planner selects the request; the executor performs the action.

SayCan · GRAPPA · RoboCerebra · SmolVLA

Task schematic: Ideal/case1, report PDF pp. 7–9. Flat receives the complete instruction; Planner may replace or retain the current subgoal. The diagram illustrates a current command, not a recorded successful rollout. Planner authority is limited to semantic goals, not motor commands. SayCan and GRAPPA motivate the interface; they are not reproduced comparison arms.

What did we test?

Fixed SmolVLA executor · three planner models · 1,230 matched episodes per planner

Normal operation

Long, multi-step tasks

Less visual input

Camera information removed

Reported faults

Predicate signal changed

Impossible tasks

Missing targets or blocked actions

Segment-based evaluation: state restored at subgoal boundaries. Reported faults alter the monitor signal.

The report explains the six-segment example. Stored initial-state restoration fixes the starting scene; demonstration-state restoration later occurs at subgoal boundaries. These are different operations. Scripted grasp/drop/wrong-object faults alter the predicate stream, leaving simulator state and executor observations unchanged. This is not uninterrupted end-to-end manipulation. All precise implementation details are manuscript/report descriptions; executable code is unavailable.

Compare the same starting state

The same success rate can hide different successful episodes.

Restore the same initial state

Flat

Complete instruction

Fixed executor

Planner

Intermediate goals

Same executor

Same task, start, stress condition and control budget

Flat	Planner	Outcome
Success	Success	Both succeed
Failure	Success	Rescued
Success	Failure	Newly failed
Failure	Failure	Both fail

Initial-state pairing and later subgoal-boundary restoration are separate operations.

Explain the four paired outcomes. A changed outcome does not by itself establish a planner effect because the fixed executor is stochastic. A/A estimates this own-system variability with the same outcome definitions. Initial-state pairing is distinct from the later segment resets.

Manipulation: progress and changed outcomes

Qwen2.5-VL: +4.0 progress points · 70 rescued / 52 newly failed · net +18 completions (p = .12).

Upper block: 1,230 matched pairs, 13 conditions. Shaded rows use their own baselines.

Previously successful episodes can fail under guidance

Among the 176 episodes solved by the fixed policy:

Goal source	Successes lost	Conditional loss rate
Qwen2.5-VL	52 / 176	29.5%
Qwen3	56 / 176	31.8%
Reference goals	76 / 176	43.2%
Kimi-K2.5	77 / 176	43.8%

A/A replay loss estimate: 14.8% [7.0, 28.9]. Clustered uncertainty is broad; these rates alone do not establish a causal planner effect.

Manipulation: a paired execution demonstration

Left: Flat model (policy alone)Right: Planner-guided policy

Flat on the left · Planner on the right · original playback speed

Kimi-K2.5 planner · shared execution clock · fixed policy in both branches.

Author-supplied paired-execution demonstration; model identity visible in the footage.

Almost finished is still unfinished

Policy aloneWith planner

Final state under the same control budget

One matched episode

6 / 6Policy alone

5 / 6With planner

The guided policy gets close,
but the task is incomplete.

Kimi-K2.5 · Ideal/case5/seed0
Different from the Qwen headline experiment.

Count completed tasks as well as partial progress.

This is Ideal/case5 under Kimi-K2.5, not the case1 teaching scene or the Qwen2.5 aggregate. Frames share a control clock. Nearly all learned-planner-induced failures in the paper end in timeout, but these frames alone do not establish a specific causal mechanism.

Manipulation study: findings

More progress did not clearly translate into more finished tasks.
Guidance helped intermediate steps, but the overall completion benefit remained uncertain.

Guidance could help one attempt and disrupt another.
Some tasks that succeeded with the policy alone failed after planner intervention.

Useful guidance still depended on execution.
Reference subgoals did not ensure completion; extra time did not improve Qwen3 completion in the tested comparison.

Qwen2.5: +4.0 progress points; +18 completions, p=.12. Goal sources lost 29.5–43.8% of fixed-policy successes. A 50% larger budget left Qwen3 completion unchanged; do not generalize this to every planner.

Planner pipeline: wheeled-robot navigation

Second study: object search with a fixed detector and Nav2 navigation stack.

Wheeled robot: task and comparison

Can planner guidance help the robot reach the requested object when the baseline cannot construct or retain a usable target?

Flat system

Fixed search, detection and approach sequence.

The robot must reach the designated object, not just another object of the same class.

Planner-enabled system

Qwen3-VL selects navigation skills and uses target context and memory.

Same starting pose and heading; five scenes, five trials per scene and system.

Fixed detector and Nav2 executor. This evaluates the broader planner-enabled stack, including memory and interpretation.

Mobile navigation: scene-level completion

System	1 Visible and reachable	2 Fixed scan	3 Outside the map	4 Around the corner	5 Implicit description	Total
Flat	5/5	5/5	1/5	2/5	0/5	13/25
Planner	5/5	5/5	5/5	4/5	5/5	24/25

Five reused scenes × five trials; manual, unblinded scoring. Comparison includes memory and semantic interpretation.

Guidance helped recover a usable target

Target localized → approach blocked
Reposition → retry the remembered target

13/25Flat

24/25Planner

Target	Flat	Planner
Already usable	10/10	10/10
Needs memory or interpretation	3/15	14/15

Success: reached the designated object.

Benefit concentrated where the baseline lacked a usable target.

Physical mobile robot · detector + Nav2 · five scenes × five trials · unblinded scoring

The planner reads the map, chooses a corridor, repositions and retries the remembered target. This broader stack comparison includes memory and semantic interpretation; it is not a pure goal-selection ablation. Five trials share each scene. Counts are manuscript-reported manual adjudications, not independently reconstructed from logs. Use descriptive scene labels because IDs differ across drafts.

Mobile navigation: observation and execution

First half: supplied mobile-robot demonstration · 2× speed

Second half: continued demonstration · 2× speed

Interpret the requested target, retain useful observations, and navigate using the detector and Nav2 stack.

Author-supplied physical mobile-robot demonstration; 2× playback speed.

Wheeled-robot study: findings

The planner helped when the target was hard to identify or remember.
Context and memory helped continue the search and approach.

Simple cases already worked with the baseline.
When the target was visible and usable, extra guidance added little.

The benefit came from the whole system.
Planning, memory and interpretation worked with the fixed detector and navigator.

The evidence is limited to the tested scenes.
Broader reliability still needs evaluation.

Planner 24/25, baseline 13/25. Already usable targets: both 10/10. Other scenes: 14/15 versus 3/15. Five reused scenes; manual unblinded scoring. This broader stack comparison does not isolate goal selection.

Part 2 · Learning for a new embodiment

1. 01Planning above a fixed controller
2. 02Learning for a new embodiment
3. 03What comes next

How can a robot learn new actions from its attempts—and keep its earlier skills?

Until now, we changed the instructions given to a fixed controller. Now we change the controller through training. The robot tries a task in simulation and receives a numerical reward for success or useful progress. That reward is what we mean by training feedback. We first test reward design on a conventional arm, then adapt the approach to a cable robot.

From the base policy to our reward experiment

OpenVLA-OFT

Images + state + instruction → robot actions
Our starting policy in LIBERO

Starting action policy

SimpleVLA-RL

Collect simulated attempts;
update the policy using rewards

Rollout-and-update framework

PolicyImages + state + instruction

→

RolloutsAttempts at the task

→

RewardScore progress or success

→

RL updateImprove the policy

OpenVLA-OFT gives us the policy; we test how reward design changes its learning.

Training feedback = numerical reward · CDPR uses a different policy: SmolVLA

Kim et al., OpenVLA-OFT (2025) · Li et al., SimpleVLA-RL (2025)

OFT studies action decoding, representation and fine-tuning. SimpleVLA-RL supplies the VLA rollout-and-update framework applied to OpenVLA-OFT. Our experiment studies reward feedback on the conventional LIBERO serial-arm interface; it is not a full replication of either paper’s headline scores. LoRA adapts the policy. The later CDPR experiment uses SmolVLA and a different action mapping.

OpenVLA-OFT: the policy we build on

Inputs

Camera observations, language instruction and robot state.

Action prediction

Continuous actions predicted in chunks with parallel decoding.

Optimized fine-tuning

The OFT recipe studies action representation, decoding and learning objectives.

Our use

An OpenVLA-OFT-based policy, adapted with LoRA in the LIBERO reward study.

Kim, Finn & Liang (2025), Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. https://arxiv.org/abs/2502.19645

SimpleVLA-RL: learning from simulated rollouts

Sample trajectories, score outcomes, and update the policy from relative performance.

Li et al. (2025), SimpleVLA-RL, Figure 2. https://arxiv.org/abs/2509.09674

LIBERO reward comparison

Published starting point

OpenVLA-OFT policy
SimpleVLA-RL rollout-and-update recipe
LoRA fine-tuning in simulation

Ten LIBERO-Spatial manipulation tasks.

Our reward experiment

Same 61.0% initialization
160 updates: terminal-only and dense + phase
Terminal verifier coefficient: 5

300 validation rollouts per endpoint.

Train the manipulator policy; vary reward feedback; keep the native success test.

Manuscript sections 05b and 06e. Following the baseline framework is supported; reproducing all headline results of the original paper is not claimed. These are simulated conventional serial-arm tasks. WX200 physical evaluation remains ongoing.

Manuscript source (reviewed 29 October 2026): https://github.com/damurka5/RL_VLA_Bootstrapping/tree/c19366e134f620ce9453bccee94c3d2d588bfd34/ral_submission

Reward new progress, not repeated motion

ReachGraspLiftMovePlace

Terminal only

Final outcome

Phase feedback

Completed subgoals

Dense + phase

New progress in a phase

Within one phase

Progress0.20.60.30.60.8

New credit+0.2+0.400+0.2

No extra progress credit for returning to an old maximum.

Our LIBERO reward method. Unweighted illustrative increments; progress baseline resets at each new phase. Success verifier unchanged.

Terminal-only reward distinguishes complete success. Phase feedback marks subgoals; dense feedback also credits a new maximum within the active phase. The progress baseline resets when the phase changes. The table is an unweighted conceptual example, not recorded reward telemetry. Bounded shaping is added to terminal reward, while the native verifier determines success. This is different from the later CDPR milestone-credit rule. Source: manuscript sections/04c_libero_dense_rewards.tex; contributions of Kamil Nuriev and Egor Novikov.

Dense + phase feedback improved the endpoint

LIBERO-Spatial · OpenVLA-OFT-based policy · 160 updates each

020406080100%

Terminal only

61.7%

Dense + phase

65.3%

Common initialization: 61.0% · terminal reward weight: 5
300 validation rollouts per endpoint

+3.6 percentage points between final endpoints

Descriptive endpoint gain; independent training-seed uncertainty is not established.

Terminal-only reaches 61.7%, dense 65.3%, difference 3.6 percentage points. Dense rises 4.3 points from the shared 61.0% initialization; these are different contrasts. Each endpoint uses 300 validation rollouts. Independent-seed significance and general sample efficiency are not established. Low-start and unequal-budget conditions remain backup material.

Manuscript source (reviewed 29 October 2026): https://github.com/damurka5/RL_VLA_Bootstrapping/tree/c19366e134f620ce9453bccee94c3d2d588bfd34/ral_submission

Intermediate feedback: benefit and design cost

Feedback	Updates	Success
Terminal only	160	61.7%
Dense + phase	160	65.3%
Phase only	110	62.3%

Dense shaping rewards new progress: Δp_t = max(p_t − p_best, 0).

Dense + phase reached a higher endpoint here. Its progress functions require task-specific design; binary milestones need completion predicates.

Terminal coefficient: 5 for terminal-only and dense + phase; 1 for phase-only. Phase-only also uses fewer updates.

Current manuscript Table LIBERO, Methods dense reward section and Results. Terminal-only and dense+phase: coefficient 5, 160 updates; phase-only: coefficient 1, 110 updates. All initialize at 61.0%, 300 validation rollouts per endpoint. The tracker updates its previous maximum within the active phase, resets the baseline on phase transition, and combines bounded weighted progress increments, phase bonuses and terminal feedback. Native success remains fixed. Detailed shaping is confined to initial simple-skill acquisition in the CDPR schedule; full-task composition uses binary milestones. This is the design motivation, not a matched ablation proving equal efficiency of milestones.

A cable robot changes the control problem

Cable-driven parallel robot (CDPR) · simulation

“Put the apple into the bowl.”

Five policy channels

Translation: x, y, z
Yaw
Gripper

Fixed inverse kinematics maps
Cartesian targets to cable lengths.

Empty gripper→Grasp + lift→Carry + place

The cables support the end effector; the five labels are policy control channels, not five cable motors. CDPR experiments use MuJoCo Warp/MJ-Lab, four object categories and plate/bowl destinations. The image is the overview camera of an earlier selected successful rollout. This introduces the mechanism, not physical transfer or the current checkpoint score.

Bootstrapping control on a new robot

Acquire → consolidate → compose. Pretrained SmolVLA + a newly initialized residual.

Latest manuscript Figure 1, commit c1dfd2c29474b018e664592b26d2c3e713a4a96b (29 October 2026). The figure’s “from scratch” refers to acquiring target-robot skills, not training the VLA backbone from scratch or a demonstration-free lineage. Dense shaping is limited to initial simple skills. Skill-specific resets differ from empty-gripper full-task starts.

SmolVLA with a learned CDPR correction

Latest manuscript architecture, Methods. Frozen visual projection: 512 features plus six proprioceptive values; eight slots × five channels give a 40-dimensional prior. The five action channels are x/y/z, yaw and gripper. This architecture enables action-interface adaptation; the paper does not isolate an architecture-only gain.

CDPR: learn skills, retain them, then combine them

Learn with RLTry a simple skill

→

Check retentionTest earlier skills

→

SIL if neededRehearse successes

→

Continue RLLearn the next skill

Learn the parts

Approach, pickup and placement have separate instructions and starting states. Dense rewards guide initial progress; binary success rewards refine each skill.

Then learn the whole task

SIL means self-imitation learning. It retains earlier skills using successful actions. Before full-task RL, fit scripted-oracle and assisted complete-task trajectories. Then train from empty-gripper starts.

CDPR stage transitions were validation-guided and manual. Full-task RL and evaluation use unassisted starts.

Example: first approach the apple, then pick it up. Placement practice may start with the apple already held so the robot can learn carrying and releasing it onto a plate. After each skill, test earlier skills. If reaching deteriorates, balance earlier reaching successes with current placement successes, fit SIL, validate recovery, then continue RL. Finally learn the whole instruction from an empty gripper with binary milestones. This is Algorithm 1; SIL is conditional on degradation, not mandatory after every update. CDPR transitions were validation-guided and manual. Earlier assisted experience is inherited.

RL: improve actions using rewards

Try the task → score the attempts → update the policy.

$$\mathcal L_{\mathrm{RL}}=\mathcal L_{\mathrm{clip}}-\eta\mathcal H$$

Small policy updates

PPO-style clipping reduces the incentive for large action-likelihood changes; it does not impose a hard bound on the update.

L_RL is the training loss.
L_clip is the clipped policy loss.

Compare attempts

Group-relative credit (GRPO-style) compares attempts from the same scene.

H encourages exploration.
η controls its strength.

Manuscript Methods: clipped group-relative loss with separate upper/lower bounds, entropy regularization and DAPO-style filtering during skill learning. PPO-style refers to the clipped surrogate; credit is group-relative rather than vanilla PPO. No new loss is claimed. For the final continuation, VLA/LoRA are frozen and residual/exploration parameters train. Stored pre-clipping samples are scored conditional on the same episode offset. Keep those likelihood details out of the oral explanation.

Learning a new skill can erase an old one

Reaching success

Before learning to carry and place

64.1%

After learning to carry and place

5.3%

What happened?

Placement practice starts with the object already held.

The policy learns transport and release, but its earlier reaching skill deteriorates.

Learning the latest skill can overwrite useful earlier behavior.

Isolated reaching test · three evaluation rounds per condition.

Mean reaching success falls from 64.1% to 5.3% after placement RL, over three evaluation rounds per condition. These are isolated skill tests, not complete-task completion. The separate recovery experiment belongs in the next slide notes; no recovery table is shown here.

Manuscript source (reviewed 29 October 2026): https://github.com/damurka5/RL_VLA_Bootstrapping/tree/c19366e134f620ce9453bccee94c3d2d588bfd34/ral_submission

SIL: recover skills by imitating successful actions

SIL means self-imitation learning: fit the policy to its own successful actions.

$$\mathcal L_{\mathrm{SIL}}=\operatorname{mean}\!\left[(\text{predicted action}-\text{recorded successful action})^2\right]$$

Balance successful actions by instruction, fit the policy, check retention, then resume RL.

Separate recovery evaluation · reaching: 3 rounds; placement: 2 rounds per condition.

Skill	Before SIL	After SIL
Reaching	8.0%	31.1%
Plate placement	50.0%	52.3%
Bowl placement	23.05%	27.65%

Partial recovery; separate from the preceding forgetting test.

This is an MSE-style action regression objective, not a new representation-learning loss. The manuscript uses the expectation of squared vector error, balances instructions, masks unused action slots and refreshes policy-dependent features. Summing versus averaging action coordinates changes a constant scale. Separate recovery experiment: reaching 8.0→31.1%, plate 50.0→52.3%, bowl 23.05→27.65%. It does not isolate SIL contribution to final full-task performance, and the inherited lineage includes earlier assistance.

Manuscript source (reviewed 29 October 2026): https://github.com/damurka5/RL_VLA_Bootstrapping/tree/c19366e134f620ce9453bccee94c3d2d588bfd34/ral_submission

Full-task RL: keep credit for a successful pickup

After consolidation and oracle/assisted fitting: full-task RL · empty-gripper start

Approach ✓

Milestone achieved

→

Held pickup ✓

Milestone achieved

→

Placement ×

Task incomplete

Keep local credit

Approach and pickup retain their
milestone-specific learning signal.

Train the remaining step

Suppress the negative completion bonus
on achieved approach/pickup milestones.

Preserve useful partial behavior while learning to finish.

Milestones assign training credit; evaluation uses no external skill switching.

Milestone outcomes are compared within rollout groups from the same scene. The negative completion bonus is suppressed on already achieved approach/pickup milestones; placement receives completion credit once. Relative advantage still depends on the group, so achieved actions are not guaranteed positive advantage. This protects a credit component, not all shared parameters. No external stage controller switches between specialists during final evaluation.

Manuscript source (reviewed 29 October 2026): https://github.com/damurka5/RL_VLA_Bootstrapping/tree/c19366e134f620ce9453bccee94c3d2d588bfd34/ral_submission

CDPR: 70.61% strict completion in simulation

Development validation: 723 / 1,024 · 512 plate + 512 bowl attempts

Marginal outcome rates, all attempts

Physical grasp

92.48%

Held lift

88.28%

Strict completion

70.61%

Zero-to-100% scale.
One training lineage.
Development validation.

Selected historical successes

Apple into plate

Apple into bowl

Historical examples; separate from the 70.61% endpoint.

Transport and release remain the main bottleneck.

Strict success: grasp, held lift, intentional correct release; no carry slip or wrong placement. All attempts include numerical failures.

Current controller: 723/1,024 strict successes, 512 plate + 512 bowl attempts, one training lineage. Physical grasp 92.48%, held lift 88.28%. Only 79.98% of held-lift attempts reach strict completion, so transport and release remain limiting. All attempts, including numerical failures, stay in the denominator. Event flags overlap.

Video provenance: apple_into_bowl.mp4, scene_8b7cc629142fc21b, September 17 checkpoint step_28309431, its own evaluation 57/256 (22.3%). This selected success illustrates the task, not the later 70.61% evaluation. Starts empty, one instruction, no teacher or external stage switch. Source: current setup/results/CDPR table and media episode JSON.

Manuscript source (reviewed 29 October 2026): https://github.com/damurka5/RL_VLA_Bootstrapping/tree/c19366e134f620ce9453bccee94c3d2d588bfd34/ral_submission

Full-task completion across destination types

Outcome	All attempts
Strict completion	723/1,024 · 70.61%
Plate completion	393/512 · 76.76%
Bowl completion	330/512 · 64.45%
Carry slip	16.70%
Wrong placement	14.16%

Development validation used during training. Empty-gripper starts, one instruction, no teacher assistance. All attempts retained; event flags can overlap.

Manuscript source (reviewed 29 October 2026): https://github.com/damurka5/RL_VLA_Bootstrapping/tree/c19366e134f620ce9453bccee94c3d2d588bfd34/ral_submission

CDPR adaptation: procedure and observed outcome

Three problems

New action interface
Too few successful complete rollouts
Earlier skills degrade during new training

Three responses

SmolVLA prior + learned residual
Simple skills first; binary milestones later
Balanced self-imitation + oracle/assisted fitting

The complete pipeline produced a policy that grasps, carries and places under one instruction.

Partial skill recovery and 70.61% development-validation completion. Each ingredient’s independent contribution remains to be tested.

The transferable experience is the training schedule: acquire useful behavior, protect it during sequential learning, then compose it. The evidence supports the combined route, not a causal ablation of every ingredient. One training lineage; development panel used during training; earlier assisted experience inherited. No hardware success claim. Source: latest manuscript Methods, Experiments, Scope of the Evidence and Conclusion.

Manuscript source (reviewed 29 October 2026): https://github.com/damurka5/RL_VLA_Bootstrapping/tree/c19366e134f620ce9453bccee94c3d2d588bfd34/ral_submission

Part 3 · What comes next

1. 01Planning above a fixed controller
2. 02Learning for a new embodiment
3. 03What comes next

What should the robot predict—and when should it ask for help?

What should the robot predict—and when should it ask for help?

Physical understanding and decisions under uncertainty

Planned VLA work

01

Predict physical consequences

PINN / iPINN and JEPA: test contact and task-outcome predictions.

02

Estimate when the model may be wrong

Calibrate error estimates to trigger replanning, verification or human help.

03

Choose useful experience

Test selective simulation and replay at matched compute.
Engineering-control cartography motivates a VLA extension.

Near term: independent evaluation and physical deployment tests.

These are future directions, not an implemented unified architecture or validated VLA gains. PINN/iPINN denotes the user’s physics-informed direction; JEPA is a predictive representation approach, not intrinsically a physical constraint. Calibrated uncertainty should be judged by consequential errors and the cost of intervention. Off-policy DQN replay does not directly validate replay for GRPO.

What the experiments establish

Planning changes both successes and failures.
Count rescued and lost successes alongside the executor’s own variability.

Adaptation requires learning and retaining skills.
The combined CDPR pipeline reached 70.61% strict success in simulation development validation.

The next checks are independent evaluation and hardware tests.
Use matched component comparisons to establish which training stages contribute.

Better guidance can change both successes and failures, so we measure both. Learning for a new robot also requires retaining earlier skills while composing the whole task. Our results support the evaluated settings. Independent evaluation, matched ablations and physical deployment tests come next. CDPR is one training lineage with oracle/assisted fitting before unassisted full-task RL.

References: policies, planning and reinforcement learning

Kim, Finn & Liang (2025). Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success.

Li et al. (2025). SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning.

Shukor et al. (2025). SmolVLA.

Xiao et al. (2025). Probe, Learn, Distill.

Study manuscripts: planner evaluation and staged VLA adaptation.

Bootstrapping Vision-Language-Action Models for New Robot Embodiments with Staged Reinforcement Learning — manuscript draft

References: related workshop publications

Salloum, Mifsud Bonici, Ibrahim, Osinenko & Kornaev (2026). Diversity-Aware Adaptive Collocation for Physics-Informed Neural Networks via Sparse QUBO Optimization and Hybrid Coresets.
AI&PDE, ICLR 2026 Workshop.

Degtiarev, Ryabkova & Kornaev (2026). LLM-Driven Loss Balancing for Physics-Informed Neural Networks.
AI&PDE, ICLR 2026 Workshop.

Nurtdinov, Korshuk, Kornaev & Maloletov (2025). Evaluation of a Robust Control System in Real-World Cable-Driven Parallel Robots.
ICLR 2025 Workshop: Towards Agentic AI for Science.

Related publications; workshop venues identified explicitly.