APR presentation Sergei Satsevich new v2 (1) (2)

1. Enhancement of robotic manipulation learning via stabilization guarantees

Annual Progress Review:
Enhancement of robotic manipulation learning
via stabilization guarantees
Sergei Satsevich
Year of study: 3
Supervisor: Associate Professor Pavel Osinenko
Doctoral program: Engineering Systems
Date (13-10-2026)

2. Research background and problem statement

Research background
•Vision–Language–Action (VLA) models learn manipulation skills from demonstrations.
•Residual RL can improve learned policies, but reward optimization alone does not establish safety guarantees.
Problem statement
Develop an RL-based adaptation methods for VLA policies that improves pick-and-place reliability while guaranteeing avoidance of
predefined forbidden regions, under explicit model and uncertainty assumptions.
Planned experimental setup
Unitree Go2 + LeRobot-compatible arm: collect objects from the floor into an onboard basket, avoiding collisions with the robot body
and onboard equipment.
2

3. Methods Overview

Simulation and learning pipeline
•Isaac Lab / Isaac Sim: Go2-mounted LeRobot-compatible arm, parallel RL
environments, domain randomization.
•Demonstrations: teleoperation data collection; optional augmentation with Isaac
Lab Mimic.
•VLA baseline: fine-tune SmolVLA on pick-and-place demonstrations.
•Proposed adaptation: bounded residual RL corrections followed by a safety filter.
Level 1 — Constrained reaching
Move the arm to a target pose without object contact.
•Define the goal set and forbidden regions around the robot body and equipment.
•Develop a fallback controller; investigate CALF/Wrapper goal-reaching guarantees
under explicit assumptions.
•Evaluate constraint enforcement and its compatibility with goal-reaching.
Level 2 — Pick-and-place
Collect objects from the floor into an onboard basket, initially with Go2 standing still.
•Compare VLA, VLA + safety filter, and VLA + residual RL + safety filter.
•Test variations in base pose, object position, observations, and dynamics.
•Measure task success, constraint violations, execution time, and safety interventions.
Validation roadmap: Simulation proof-of-concept → robustness evaluation → real
Go2 + arm experiments.
3

4. Current Results and Conclusions

•Completed during the current reporting period
•Reviewed CALF, Multi-CALF, and CALF-Wrapper formulations.
•Designed and manufactured a custom LeRobot mounting system for
Unitree Go2.
•Assembled the physical platform: Unitree Go2 + LeRobot manipulator +
onboard collection bin.
•Developed a quadruped teleoperation setup.
•Defined a VLA-guided object grasping scenario with Whole-Body Control.
•Established four methods for comparison: fixed posture, classical WBC, RL,
and RL + CALF.
•Selected Isaac Lab for simulation, training, and benchmarking.
•Current implementation stage
•Hardware platform: assembled; system integration and calibration in
progress.
•Isaac Lab environment: under development.
•VLA grasping: dataset collection and policy training planned.
•Whole-Body RL: experimental setup and reward function being defined.
•CALF-Wrapper integration: planned after baseline RL and fallback-
4

5. 3D model and Simulation

3D-model in Siemens NX
Lerobot is moved forward on the robot
bachward in order to easier reach objects
Simulation of the setup for RL in Isaac Lab
5

6. Experimental setup

Teleoperation
system
Lerobot
Box
Quadrupped
robot
Lidar &
Onboard PC
Camera
6

7. Experiment stages

Reaching an object
Returning in base pose with
the object
Delivering the
object into the box
7

8. Teleoperation experiment

8

9.

Planned paper:
VLA-Aware Whole-Body Control for Floor Object Grasping by a Quadruped Robot
Idea: keep the same VLA grasping policy fixed, and compare different body-control strategies to see how much better body positioning improves
floor-object grasp success.
Scenario
Object on
the floor
Whole-Body Controller
chooses body pose
(height / pitch / foot
adjustment)
Compared methods
Camera detects
the object
Robot approaches
the object
VLA attempts grasp
with open gripper
Grasp success /
failure
The robot first sees the object, then adjusts its body posture, and only after that tries to grasp it.
Main metrics
• Grasp success rate
• Reachable grasp area
• Positioning time
• Stability / safety violations
Planned publication:
IEEE ACCESS
Scientific novelty
Method
Body control
Grasping
Baseline 1
Fixed posture
VLA
Baseline 2
Classical WBC / IK
VLA
Method 1
Whole-Body RL
VLA
Method 2
RL + CALF
VLA
Grasp success depends not only on a
"smart arm", but also on how well the robot
positions its whole body.
• Evaluate VLA together with whole-body control, not in isolation
• Learn body posture that maximizes the success of the same fixed VLA policy
• Add RL with guarantees (CALF) for more predictable and safer posture regulation
9

10. Planned papers:

1. Grasp-Preserving Body–Arm Coordination for Quadrupedal Object Collection
After a VLA-guided grasp, the robot returns from a crouched posture to standing and readjusts its feet. RL
coordinates body and arm movements to prevent object loss during this transition. We compare classical control,
RL, and RL + CALF by grasp retention, recovery time, and motion smoothness. Formal guarantees are investigated
for convergence to the target posture; object retention is evaluated experimentally.
2. Progress-Guided Recovery from VLA Manipulation Failures
A progress estimator detects when a VLA grasp attempt stalls or fails. RL selects a recovery action: retry the grasp,
adjust body posture, or retract the arm and restart. CALF is investigated for convergence to a designated reset
configuration under specified assumptions. We compare this approach against unassisted VLA and rule-based
retries using overall success rate, completion time, and operator interventions.
3. Reliable Object Transfer and Placement into an Onboard Bin
After grasping an object, the robot must transport it to its onboard bin and release it inside without dropping it
prematurely. We compare predefined trajectories, RL, RL + CALF, and VLA by placement success, object-loss rate,
transfer time, and training effort. CALF-based guarantees concern reaching a suitable release configuration;
successful placement is assessed experimentally.
10

11. Further Steps and Timeline

1st publication
2025
2026
PHD defence
2nd publication
2027
Today
Literature
review
2028
PhD Courses (18 + 12 + 3 ETCS)
RL with Formal Guarantees: CALF and Universal Policy Wrapper
Data & Simulation Stack
Isaac Lab Environments
Data Pipeline
Software
development
Guarantee Layer
DR & Metrics
Development of user-friendly UI
Experiments
and system
validation
IL/RL Baselines on Manipulation Tasks
Guarantee Verification: Goal-Reaching & Stability
Final system evaluation

12. Publications

Published and accepted papers:
1.
2.
3.
4.
5.
6.
HyperSurf: Quadruped Robot Leg Capable of Surface Recognition with GRU and Real-to-Sim Transferring Sergei Satsevich, Artem
Bazhenov, Elizaveta Pestova, Yaroslav Savotin, Danil Belov, Artem Erhov, Batyr Khabibullin, Vyacheslav Kovalev, Aleksey Fedoseev,
and Dzmitry Tsetserukou 2024 Sarawak, Malaysia, 6-10 Oct., 2024, pp. 2625-2630. (Scopus and WoS, H-index (SJR)=66)
10.1109/SMC 54092.2024.1 0831295
Optimizing energy consumption for legged robot by adapting equilibrium position and stiffness of a parallel torsion spring Danil Belov,
Artem Erhov, Farit Khabibullin, Elizaveta Pestova, Sergei Satsevich, Ilya Osokin, Pavel Osinenko and Dzmitry Tsetserukou 2024
Proc. IEEE Int. Conf. on Robotics and Biomimetics (ROBIO 2024), Bangkok, Thailand, Dec. 10-14, 2024
MissionGPT: Mission Planner for Mobile Robot based on Robotics Transformer Model V. Berman, A. Bazhenov, S. Satsevich, and D.
Tsetserukou 2024 Proc. IEEE Int. Conf. Foundation and Large Language Models (FLLM 2024), Dubai, UAE, 26-29 Nov., 2024, pp.
201-206. 10.1109/FLLM 63129.2024.1 0852423
ViewVR: Visual Feedback Modes to Achieve Quality of VRbased Telemanipulation A. Erkhov, A. Bazhenov, S. Satsevich, D. Belov, F.
Khabibullin, S. Egorov, M. Gromakov, M. Altamirano Cabrera, and D. Tsetserukou 2025 Proc. IEEE/ACM Int. Conf. on Human Robot
Interaction (HRI 2025), Melbourne, Australia, March 4-6, 2025, pp. 1294-1298. (Core2023 A, Scopus, WoS, H-index (SJR)=70)
10.5555/3721 488.3721666
Prometheus: Universal, Open-Source Mocap-Based TeleoperationSystem with Force Feedback for Dataset Collection in Robot
Learning Sergei Satsevich*, Artem Bazhenov*, Sergei Egorov, Artem Erkhov, Maxim Gromakov, Aleksey Fedoseev,and Dzmitry
Tsetserukou 2025 IEEE ROBIO Dec. 3-7, 2025, Chengdu, China
Echo: An Open-Source, Low-Cost Teleoperation System with Force Feedback for Dataset Collection in Robot Learning
Artem Bazhenov*, Sergei Satsevich*, Artem Erkhov, Dzmitry Tsetserukou 2026 IEEE 22nd International Conference on
Automation Science and Engineering August 17 – 21, 2026, Shenyang, China
12

13. Courses

Planned courses
Credits
Research Methodology for Engineering Systems
3
Reinforcement Learning
6
History and Philosophy of Science
6
English. Candidate Exam
3
Leadership for Innovators
3
Pedagogical Experience
3
Thesis Proposal
6
Term 10
Term 9
Term 11
• Leadership for Innovators
• History & Philosophy
Term 12
Research Methodology for Engineering
Systems
Reinforcement Learning
Term 13 - 14
English. Candidate Exam
Term 15
• Pedagogical Experience
13