DReS Robot Cell Simulator - AI Agent Finds the Collisions of a Robot Station with One Prompt (YouTube, English)
DReS Robot Cell Simulator - KI-Agent findet die Kollisionen einer Roboterstation mit einem Prompt (YouTube, Deutsch)

AI Agent window API / CLI

Find all possible collisions of the open station: baseline run, work zones, check interlocks, then write and run a robustness test plan with Live 3D. Give me the report.

The second tab next to the Message Window (bottom, AI icon) is a chat with an AI agent. You write what the agent should do, e.g. "Run the simulation of the station and find all possible collisions"; a language model plans the steps and uses the commands of the simulator like the buttons, without dialogs. The chat shows your messages, the answers of the agent and every command with its result.

The buttons New chat, settings, Send and Stop have one width, so the right edge of the conversation and of the message field is the same.

The agent works: the state line shows the step and the progress of the simulation (Claude Code is working... (step 4) - simulation running: cycle 1 / 1); every command of the agent with its result

Answer live: with all three ways of connection (A, B, C, see Configuration) the answer appears word by word while the model writes it. When a block of text is complete, it is shown formatted (bold, lists, code). Reasoning blocks of the model (thinking) are supported: they are kept in the conversation and sent back to the model, but not shown.

Tables in the answers of the agent (Markdown tables, e.g. the cycle times of the robots) are shown with aligned columns in a fixed-width font, so the numbers stand one below the other.

A full answer: cycle time of the station, table of the cycle times of the robots (aligned columns), waits, a request race on zone 1, contacts and the programs that were not simulated; state line context 40 % (green) and the tokens of the chat

Usage in the state line: after the first answer the state line shows context NN % – how full the context window of the model was with the last request (setting Context window [tokens]) – and the tokens of this chat (in, out), optionally with (NN % of the limit) of the setting Token limit per chat. The colors are like in the window of Anthropic: green below 60 %, orange below 85 %, red from 85 %. When it is red, start a New chat – a full context makes the agent slower and less exact. The limit is only a warning, the agent does not stop.

Change of the provider or the model: a new provider or model (in the settings) always starts a new chat – the token counters start from zero. When the change is made while the agent works, the current task is finished first and the new chat starts as soon as the agent has finished.

Messages while the agent works: a message sent while the agent works waits in a queue ((queued - sent when the agent has finished; …)) and is sent when the agent has finished; Stop ends the waiting. Exception in the ways A and C (API): when the agent waits for the end of the simulation, your message ends this waiting and goes to the agent at once ((sent to the agent while it waits for the simulation)). The agent handles it – e.g. hiding the paths works during the simulation – and then waits again. The simulation goes on. In the way B (Claude Code) the message always waits in the queue. The commands in the chat (/help, /reload and the direct commands of the simulator) never wait – they run at once.

Question before the first real data: when the anonymous mode is off, the first message of a DReS session that would go to the model is held, and the chat asks (orange Question) whether the real data of the cell may be sent. Answer in the chat: YES (also JA / TAK) – send the real data (asked only once per session); ANON – switch on the anonymous mode and send the data anonymized; NO (NEIN / NIE) – cancel, nothing is sent. Another text gives Please answer YES, ANON or NO. The local commands (/help, /status …) work also while the question waits. The question can be switched off in the settings: check box Ask before the first sending of real data (YES / ANON / NO) (in the simulator on by default, see Settings).

Font size of the conversation and of the message field: Ctrl + mouse wheel, Ctrl+Plus / Ctrl+Minus or the context menu (Larger font, Smaller font), 7 … 32 pt. The size is remembered (setting RCS_Group/AiChat_FontSize).

The tabs of the bottom area (Message Window / AI Agent) have their titles at the top; the tab of the agent carries the AI icon.

Welcome: every time the plugin is opened and after New chat the chat shows a short line with a link to this documentation and the hint type /help.

The AI Agent tab after New chat: welcome line with the link to this documentation

Commands in the chat (they work locally – no request to the model, no license and no API key needed):

• /help – a short brief: what the agent does, examples of messages, the commands, the current connection (provider), the path of the log of the agent and a link to this documentation.

/help: what the agent does, examples, the commands, the connection and the log

• /reload (also reload knowledge, przeładuj wiedzę, odśwież wiedzę, Wissen neu laden) – loads the knowledge files again. The chat shows an entry with the summary (number of files, characters, tokens, missing files). The conversation stays.
• /anonymous on | off (also /anon on | off) – switches the anonymous mode on or off (a new chat starts); /mode – shows the mode and, when it is on, the key R1 = 0001R01, ….
• Direct commands of the simulator – every other message that starts with / and a letter is a command of the command API, run at once in the program: without the model and without tokens, also while the agent works (no queue). Syntax: /name key=value … (values true / false, numbers, text; text with spaces in quotes, e.g. name="R06 UP42") or /name {json}. Examples:
  • /status, /set – state of the simulator, settings;
  • /run folge=125 loops=1 sps=false – starts the simulation;
  • /view programPaths=false recordedPath=false – hides the paths in the 3D view;
  • /zones_view show=true overlapOnly=false – shows the work zones including the green ones (areas of one robot only);
  • /robot {"robot":"R06","goto":"UP42/P17"} – the same in JSON.
  The JSON answer appears in the conversation (an error in red); the log of the agent gets an entry direct with the command and the answer. Commands that change something need the license API / CLI, like in the command API. /ai_chat is not available in the chat. The list of the commands and their parameters: command API.

Direct command /set: the simulator answers at once (JSON) – without the model, no tokens

Tools of the agent: rcs_command – any command of the command API, run inside the simulator (no pipe, no windows); wait_until_idle – waits until loading, preparation, simulation and robustness test are over; read_doc – reads the pages doc\rcs_*.html of this documentation (e.g. rcs_api before it uses a command it does not know). Steps per message in the settings limits the number of commands for one message. A file the user names ("load my example", "the last simulation") the agent finds in the list of the recent simulations (command recent_files) – it does not search the disk.

Instructions of the agent: <installation>\template\RCS\AIAgent\rcs_agent_instructions.md (the default knowledge file; can be adapted or extended by your own files). The standard procedure find all possible collisions:

1. state of the simulator (load the station if needed),
2. tool models on the robots (only when the user asks: tool_model, check with robot goto in a weld point, see Tools on the robots),
3. HOME first: before every simulation of a program / Folge all robots go to HOME with Folge 125 (run folge=125 loops=1 sps=false, wait for the end); only then the real run – otherwise the cycle times start from random positions,
4. learning run: 1 normal cycle, cycle time and order of the robots,
5. request races and the Interlock priority blocks added to the SPS program,
6. work zones and Check interlocks (MISSING, LATE, EARLY, ...),
7. robustness test with Live 3D and a plan made by the agent for the risky places,
8. round 2 when places are still not checked,
9. report: cycle time, races, every collision place with its cause, what could not be tested.

The agent always answers every message – also when something cannot be done: then it says why (e.g. the work zones are not calculated yet). The instructions contain a dictionary of the display, so the agent uses the same words as the 3D view: green zones = area of one robot only, yellow = common area with an interlock, red = common area without an interlock; paths in the colors of the interlocks, labels. The zones must be calculated first (command zones) – not while the simulation runs.

License: the window needs the additional license API / CLI (L_248_API). Without it the messages are rejected and the group of the connection in the settings is locked (/help and /reload still work; the direct commands follow the rules of the command API).

Anonymous mode

In the anonymous mode the model does not see the names of the cell: the simulator replaces them by identifiers before a text goes to the model and puts the real names back in what comes from the model. The study runs as usual – the agent simulates, checks and reports with R1, R2 … instead of the robot names.

Switching the mode starts a new chat. While the mode is on, the state line shows anonymous in green.

Both directions: the answers of the commands and your messages go to the model with the identifiers; the commands of the model (e.g. {"cmd":"robot","robot":"R2","goto":"UP42/P5"} or {"cmd":"load","file":"<file 1>"}) come back to the simulator with the real names; the answer of the agent is shown to you with the real names. The lines of the tools in the chat show what the model saw.

All connections: Anthropic API (A), Claude Code (B) and VW LLMaaS / OpenAI-compatible API (C). With Claude Code the MCP server VeRa gets the environment variable DRES_AI_ANON=1 and adds "anon": true to every command of the simulator; in this mode Claude Code may not read files (no Read, Grep, Glob – the files contain the real names).

Speed: the simulation itself is not changed and not slower – only the text of the commands and answers is translated (fractions of a millisecond). The key (real name ↔ identifier) stays only in the memory of the simulator.

When the mode is off, the chat asks before the first real data are sent (YES / ANON / NO, see Question before the first real data).

Configuration step by step

Your own access to the AI model: the agent uses your account or API key (Anthropic, an OpenAI-compatible service such as VW LLMaaS / Azure OpenAI, your own claude.ai account) or a local model. The usage of the model is paid directly to your AI provider under your own contract – DReS adds no AI fees and no usage limits of its own.

Choose one of the three ways to connect the agent to a language model. For all of them the license API / CLI is needed.

A. Anthropic API (Claude) – with an API key

1. Open console.anthropic.com and create an account. In Settings > Billing buy credit: you pay for what you use. The API is not part of a Claude Pro / Max plan.
2. Settings > API Keys > Create Key. Copy the key (sk-ant-...) – it is shown only once.
3. In the simulator: Tools > Settings... > AI Agent. Provider: Anthropic API (Claude), paste the key in API key / token. Model: empty = claude-sonnet-5-5; claude-opus-5-5 is more exact and costs more. OK.
4. Tab AI Agent: type /help, then your message, e.g. "Run the simulation of the station and find all possible collisions".

B. Claude Code – your own claude.ai account (Pro / Max), no API key – only for your own use

1. Install Claude Code (claude.com/claude-code).
2. Log in – before the first use and when the session has expired: the program claude.exe started by the simulator has its own login, independent of the Claude desktop application. Open a terminal (e.g. cmd or PowerShell), run claude, log in with /login (your claude.ai account) and close it with /exit. When the login is missing or has expired, the chat shows Failed to authenticate… and a red hint with these steps – log in and send the message again.
3. Python 3 must be installed – it runs the MCP server AIAgent\vera_mcp.py (found automatically, see Python below).
4. In the simulator: Tools > Settings... > AI Agent:
   • switch on Command API of the Robot Cell Simulator for the AI agent and scripts (pipe DReS_RCS_API) – Claude Code controls the simulator through the MCP server VeRa;
   • Provider: Claude Code - your claude.ai account (own use only);
   • Claude Code program and Python: filled in automatically the first time the settings are opened; check them or choose the file with the button next to the field;
   • Model: empty = the default of Claude Code (e.g. sonnet or opus); OK.
5. Tab AI Agent: write your message.

How it works: the first message of a chat starts claude -p --input-format stream-json --output-format stream-json --include-partial-messages ... with the MCP server VeRa (tools rcs and status); the next messages go to this running process. The knowledge files go in as additional instructions (--append-system-prompt-file), the documentation doc\ can be read (tool Read). The conversation goes on (--resume); New chat starts a new process and a new session. The tokens come from the limit of your plan; the state line shows the tokens used.

C. VW LLMaaS or another OpenAI-compatible API

1. Tools > Settings... > AI Agent, Provider: OpenAI-compatible API.
2. Address: the address from the LLMaaS portal (.../chat/completions); API key / token: your token.
3. Key header: Authorization (Bearer token, default) or api-key (Azure OpenAI).
4. Model: a model from the LLMaaS list, e.g. claude-sonnet-4.6 or gpt-5.1; OK.
5. Data: only Confidential or lower (see Data protection below).

Tools > Settings... > AI Agent: provider Claude Code, Claude Code program and Python with the buttons to choose the file, address, key, header, model, steps, context window, token limit, knowledge files (Add..., Remove, Default, Refresh knowledge), Documentation in the lower left corner

Settings (Tools > Settings... > AI Agent, group AI agent window (tab next to the messages) – connection to the model):

Credit and costs (way A): the remaining credit is shown only in console.anthropic.com > Billing; the usage and the costs in Console > Usage / Cost. Through the API they can be read only with an administration key of the organization (sk-ant-admin...), not with the normal API key – the simulator does not read them. In the window: the tokens of the chat in the state line and every request in the log of the agent.

The knowledge is read once – with the first message, with Refresh knowledge, /reload or OK in the settings – and stays the same between the requests. The provider can then use its prompt cache, so long instructions cost less. A changed file is used only after /reload or Refresh knowledge.

Fixed system instructions: the instructions sent to the model contain only the knowledge files – no changing data of the scene (number of robots, state). The agent reads the current state itself with the command status. So the instructions stay the same during the whole chat and the prompt cache works. After /reload or a change of the knowledge the reasoning blocks of the model (thinking) of the earlier answers are removed from the history (they are signed for the old instructions – the API would answer HTTP 400 Invalid signature in thinking block); the conversation itself stays.

Log of the agent: %TEMP%\dres_rcs_agent\ai_agent_YYYYMMDD.log (one file per day). It holds every entry of the conversation, the commands and the full results of the tools (the window shows only a short form), HTTP errors with the text of the answer, and for Claude Code the final result, its error output (stderr) and the exit code. /help shows the path. The folder %TEMP%\dres_rcs_agent also holds the working files of the provider Claude Code (mcp_vera.json – the MCP server VeRa, rcs_agent_prompt.md – the instructions of the agent); they are written again for every message.

Problems (the details are in the log of the agent):

Script / API: command ai_chat (message, newChat, stop, last, anonymous – answer busy, transcript, anonymous, consentPending), see the command API. Other ways to use AI: anonymized study for an AI assistant, the AI Agent window of the editor (same connection) and VeRa (voice control, AI assistant over MCP).